Panel Session
Brain-Mind Architectures: Module-Free, General Purpose, and Immediate Learning?



Top | Abstract | Organizers | Tentative Panelists | Email Discussions

Panel session for IJCNN 2011

Abstract

During his talk at IJCNN 2006, Bernard Widrow criticized the existing models of neural networks, driven by error back-propagation, Hebbian mechanisms or other known mechanisms, to be slow and prone to local minima.   An alternative approach is proposed by Steve Pinker, who wrote: “The mind, I claim, is not a single organ but a system of organs, which we can think of as psychological faculties or mental modules.” (p. 27, How the Mind Works).  Recently, a generative developmental network (IJCNN 2010) was claimed to be general-purpose reasoning (in the sense of finite automata), module-free, and to perform immediate learning.  In neuroscience, the Broca’s area which was believed to be associated with languages, has shown to be involved in many other activities.

This panel will bring active researchers to discuss the following array of challenging questions:

  1. Is the brain’s memory general purpose?  In which way?
  2. Does the brain have rigid modules over the life?  Why?
  3. How does the brain-mind learn?  Fast?   Is Hebbian learning sufficient for fast, immediate learning?
  4. How does the brain-mind reason?  Equivalent to a complex finite automaton?  Why yes and why not?
  5. Is it the time for connectionist approaches to consider a new picture of brain-mind?  What and Why?

This panel is co-sponsored by:
(a) IEEE CIS Autonomous Mental Development TC and its Visual Processing Task Force
(b) INNS Autonomous Learning SIG

Organizers

Juyang (John) Weng, Professor
Embodied Intelligence Laboratory
Department of Computer Science and Engineering,
Cognitive Science Program, and Neuroscience Program
Michigan State University
East Lansing, MI 48824 USA
Tel: 517-353-4388, 617-253-8024
Email: weng@cse.msu.edu
http://www.cse.msu.edu/~weng/

Asim Roy, Professor
Dept. of Information Systems
Arizona State University
Tempe, AZ 85287-4606 USA
Email: asim.roy@asu.edu
http://lifeboat.com/ex/bios.asim.roy

Tentative Panelists

Prof. Asim Roy, Arizona State University, USA, asim.roy@asu.edu
Prof. Ron Sun, Rensselaer Polytechnic Institute, USA, rsun@rpi.edu
Prof. Janusz Starzyk, Ohio University, USA, starzykj@ohio.edu
Prof. John Taylor, King’s College London, UK, john.g.taylor@kcl.ac.uk
Prof. Juyang Weng, Michigan State University, USA, weng@cse.msu.edu
Prof. Bernard Widrow, Standford University, USA, widrow@stanford.edu

Email Discussions

-------- Original Message --------
Subject: Christof Koch revives the grandmother cell idea??
Date: Wed, 22 Jun 2011 12:21:23 -0700
From: Asim Roy <ASIM.ROY@asu.edu>
To: Walter J Freeman <dfreeman@calmail.berkeley.edu>, "Dr. Bernard Widrow" <widrow@stanford.edu>, "Taylor, John" <john.g.taylor@kcl.ac.uk>, Kunihiko FUKUSHIMA <fukushima@m.ieice.org>, Nikola Kasabov <nkasabov@aut.ac.nz>, "Fredric Ham" <fmh@fit.edu>, <dfilev@ford.com>, Risto Miikkulainen <risto@cs.utexas.edu>, Jacek M Zurada <jmzura02@louisville.edu>, Carlo Francesco Morabito <morabito@unirc.it>, "Srinivasa, Narayan" <nsrinivasa@hrl.com>, Klaus Obermayer <oby@cs.tu-berlin.de>, George Lendaris <lendaris@sysc.pdx.edu>, Kenji Doya <doya@oist.jp>, Evangelia Micheli-Tzanakou <etzanako@rci.rutgers.edu>, Danil Prokhorov <dvprokhorov@gmail.com>, Jose Principe <principe@cnel.ufl.edu>, Lee Giles <giles@ist.psu.edu>, bruno apolloni <apolloni@dsi.unimi.it>, "Ronald R. Yager" <yager@panix.com>, Thomas Caudell <tpc@ece.unm.edu>, DeLiang Wang <dwang@cse.ohio-state.edu>, Richard M Golden <golden@utdallas.edu>, Marley Vellasco <marley@ele.puc-rio.br>, Amir Atiya <amir@alumni.caltech.edu>, "Sarangapani, Jagannathan" <sarangap@mst.edu>, Leif Peterson <peterson.leif@ieee.org>, Kevin Logan <klogan@macsea.com>, Ivette Luna Huamani <iluna@cose.fee.unicamp.br>, Emilio Del Moral Hernandez <emilio@lsi.usp.br>, José Alfredo F. Costa <alfredo@ufrnet.br>, Roseli Aparecida Francelin Romero <rafrance@icmc.usp.br>, "Khan M Iftekharuddin (iftekhar)" <iftekhar@memphis.edu>, <Yaochu.Jin@surrey.ac.uk>, Chiranjib Bhattacharyya <chiranjib.bhattacharyya@rediffmail.com>, Marimuthu Swami Palaniswami <palani@unimelb.edu.au>, Salim Bouzerdoum <bouzer@uow.edu.au>, <mjhealy@ece.unm.edu>, "OZAWA, Seiichi" <ozawasei@kobe-u.ac.jp>, Nikolik <nikolik@msm.nl>, <richard@udc.es>, José García <jgarcia@dtic.ua.es>, "Adel M. Alimi" <adel.alimi@ieee.org>, <spang@aut.ac.nz>, <p.angelov@lancaster.ac.uk>, <syxie@nwpu.edu.cn>, Ivan Aquino <ivaqmo@gmail.com>, "Venayagamoorthy, Ganesh K." <ganeshv@mst.edu>, Prasad Girijesh <G.Prasad@ulster.ac.uk>, Leon Reznik <lr@cs.rit.edu>, Ricardo Sanz <Ricardo.Sanz@upm.es>, Pascual Campoy <pascual.campoy@upm.es>, Tao Ban <bantao@nict.go.jp>, <rich.hammett@gatech.edu>, Wlodzislaw Duch <wduch@is.umk.pl>, <Edgar.Koerner@honda-ri.de>, <yamauchi@cs.chubu.ac.jp>, <sasi001@gannon.edu>, <priyanka2309@gmail.com>, <bailly@isir.fr>, <Nistor.Grozavu@lipn.univ-paris13.fr>, <hmeng@lincoln.ac.uk>, <parena@diees.unict.it>, <rharley@ee.gatech.edu>, <anne@dimap.ufrn.br>, <ahunter@lincoln.ac.uk>, <kappiah@lincoln.ac.uk>, <okada_s@i.kyoto-u.ac.jp>, <hahosoya@is.s.u-tokyo.ac.jp>, <fredrik.sandin@gmail.com>, <nils@siebel-research.de>, <rwindecker@aol.com>, <marco.piastra@unipv.it>, <tamas.jantvik@ltu.se>, <petia@ua.pt>, <perdi@kzoo.edu>, <dwunsch@mst.edu>, Mustapha Lebbah <mustapha.lebbah@univ-paris13.fr>, Juyang Weng <weng@cse.msu.edu>, Ravi Rao <ravirao@us.ibm.com>, Professor Ron Sun <rsun@rpi.edu>, Yan Meng <yan.meng@stevens.edu>, Minoru Asada <asada@ams.eng.osaka-u.ac.jp>, <ali.minai@uc.edu>, <leonid@hrl.harvard.edu>, Janusz Starzyk <starzykj@gmail.com>, Angelo Cangelosi <acangelosi@plymouth.ac.uk>, Giorgio Metta <pasa@dist.unige.it>, Pierre-Yves Oudeyer <pierre-yves.oudeyer@inria.fr>, <hussain@ieee.org>, <rolf.wuertz@ini.rub.de>, <francesco.gauupp@cs.man.ac.uk>, <elexujx@nus.edu.sg>, <gpazienza@sztaki.hu>, Janusz Starzyk <starzykj@bobcat.ent.ohiou.edu>, <fred@ini.phys.ethz.ch>, <bntlyb@gmail.com>, <mauro.tucci@dsea.unipi.it>, <rim@cse.snu.ac.kr>, <Damiano.Oldoni@UGent.be>, <mlovric@rsm.nl>, <p.wawrzynski@elka.pw.edu.pl>, <norifumi@tamagawa.ac.jp>, <tgibbons@css.edu>, <i.nica@surrey.ac.uk>, <harry.erwin@btinternet.com>, <anhngv102@gmail.com>, <swban@dongguk.ac.kr>, <lperl@rcn.com>, <ogchang@gmail.com>, <red51@epsam.keele.ac.uk>, <m.Lukosevicius@jacobs-university.de>, <yjchou@mail.ndhu.edu.tw>, <christian.faubel@ini.rub.de>, <jzhu@fias.uni-frankfurt.de>, Giacomo Indiveri <giacomo@ini.phys.ethz.ch>, Sylvie Renaud <sylvie.renaud@ims-bordeaux.fr>, <chicca@ini.phys.ethz.ch>, <paolo.mottoros@polito.it>, <corazza@unive.it>, <aurel@ieee.org>, <bertoni@dsi.unimi.it>, <silvia@sa.infn.it>, <masulli@disi.unige.it>, <giacco@sa.infn.it>, <mirko.esposit@gmail.com>, <re@dsi.unimi.it>, <bassis@dsi.unimi.it>, <mirko.lucchese@gmail.com>, <lorenzo.valerio@gmail.com>, <michele.scarpiniti@uniroma1.it>, <micheli@di.unipi.it>, <paolo.delgiudice@iss.infn.it>, <dlester@cs.man.ac.uk>, Alexander Rast <rasta@cs.man.ac.uk>, Rodrigo Calvo <calvo.rodrigo@gmail.com>, José García <jgarcia@dtic.ua.es>, <aal@clopinet.com>, Dileep George <dileep@vicariousinc.com>
CC: <koch@klab.caltech.edu>, "Jay McClelland" <mcclelland@stanford.edu>, "jeff bowers" <j.bowers@bristol.ac.uk>, <plaut@cmu.edu>, <ja+@cmu.edu>, <markman@psy.utexas.edu>, "Gerry Altmann" <cognition@york.ac.uk>, "Tim Rogers" <ttrogers@wisc.edu>, "Rick Cooper" <R.Cooper@bbk.ac.uk>, "Matthew Crocker" <crocker@coli.uni-sb.de>

In case you missed the recent Scientific American Mind article by Christof Koch (http://www.scientificamerican.com/article.cfm?id=being-john-malkovich ). Below is a quote from the article. He not only resurrects the notion of the grandmother cell, but also relates it to the phenomenon of consciousness. This bodes well for the theory that the brain uses both local and distributed representation.
“One hippocampal neuron responded only to photos of actress Jennifer Aniston but not to pictures of other blonde women or actresses; moreover, the cell fired in response to seven very different pictures of Jennifer Aniston. We found cells that responded to images of Mother Teresa, to cute little animals and to the Pythagorean theorem, a2 + b2 = c2….. Nobody is born with cells selective for Jennifer Aniston. Like a sculptor patiently releasing a Venus de Milo or Pietà out of blocks of marble, the learning algorithms of the brain sculpt the synaptic fields in which concept neurons are embedded. Every time you encounter a particular person or object, a similar pattern of spiking neurons is generated in higher-order cortical regions. The networks in the medial temporal lobe recognize such repeating patterns and dedicate specific neurons to them. You have concept neurons that encode family members, pets, friends, co-workers, the politicians you watch on TV, your laptop, that painting you adore.”
Christof Koch. Being John Malkovich. Scientific American Mind, March/April, 18-19, 2011.



On 6/23/11 3:40 AM, Ali Minai wrote:
I think most people accept that the brain uses both local and distributed representations in the sense that particular cells can become very specifically tuned. There is extensive work on this both in the semantic cognition literature and in the visual representation literature. Earl Miller's work on cats and dogs, for example, demonstrated this, as did the beautiful work by Tsunoda, Yamane et al on the representation of object structure.

Best

Ali


On 6/23/11 5:43 AM, Christof Koch wrote:
Correct. We have shown that the population as a whole responds very sparsely to familiar individuals, like Jennifer Aniston, and that single neurons are highly selective.
We estimate - using bayesian reasoning - that there might be up to 10^5 neurons that respond to any one specific familiar individuals, with most cells responding to more than one. The detailed math is in various papers.

C.
----
Dr. Christof Koch



On 6/23/11 7:37 AM, Asim Roy wrote:
Is it a fair to say that under distributed representation, single neurons should not be highly selective and should not respond to only one kind of stimuli? And that should be true whether it dense or sparse.

Asim



On 6/23/11 3:40 AM, Ali Minai wrote:
I think most people accept that the brain uses both local and distributed representations in the sense that particular cells can become very specifically tuned. There is extensive work on this both in the semantic cognition literature and in the visual representation literature. Earl Miller's work on cats and dogs, for example, demonstrated this, as did the beautiful work by Tsunoda, Yamane et al on the representation of object structure.

Best

Ali


On 6/23/11 5:43 AM, Christof Koch wrote:
Correct. We have shown that the population as a whole responds very sparsely to familiar individuals, like Jennifer Aniston, and that single neurons are highly selective.

We estimate - using bayesian reasoning - that there might be up to 10^5 neurons that respond to any one specific familiar individuals, with most cells responding to more than one. The detailed math is in various papers.
C.
----
Dr. Christof Koch
On 6/23/11 7:37 AM, Asim Roy wrote:
Is it a fair to say that under distributed representation, single neurons should not be highly selective and should not respond to only one kind of stimuli? And that should be true whether it dense or sparse.

Asim



On 6/23/11 7:41 AM, Christof Koch wrote:
This is an excellent and very short article by P. Foldiak on these questions.
Cheers
C.
----
Dr. Christof Koch


On 6/23/11 8:03 AM, Ali Minai wrote:
Asim

I don't think so. In fact, I would expect effective distributed representations to show high modularity (in the systems biology sense), and to follow the principles of "near decomposability" (Simon) and/or "weak linkage" (Gerhardt & Kirschner). If we've learned anything from complex biological systems, it is that they show just enough connectivity to be able to exploit it and not so much that all change becomes impossible. On this principle alone, neurons participating in good distributed representations should be very sparsely tuned and very selective.

Best

Ali


On 6/23/11 9:22 AM, Juyang Weng wrote:
I quote from Foldiak and I added my comments according to our computational brain model
(e.g., Weng, IJCNN 2010):

"The latest paper in this series, published recently in Current Biology [6],
demonstrates that many of the recorded neurons respond not only
to images of one specific item, for example ‘‘Saddam Hussein’’, but also
to the written and spoken name of the same item."

[Weng: According to our computational model, this ‘‘Saddam Hussein’’ like cell is
not unique in the brain. Thus, this is not a grandmother cell. I guess that one person has only
one grandmother. There are many many ‘‘Saddam Hussein’’ like cells,
in MTL, the frontal cortex, and the pre-motor cortex. Then, how does the brain use these cells?
An answer is in Weng IJCNN 2010]

"The auditory and visual selectivities are precisely aligned,
so that the auditory, visual textual descriptions and visual images
activating a given neuron correspond to the same real-world objects."

[Weng: I computationally explained in Weng IJCNN 2010 how many such cells can
develop fully autonomously inside the brain without requiring the teacher to
opening the skull of the brain to implant symbolic Bayesian framework.
The type of models in Tenenbaum et al. "How to Grow a Mind: Statistics, Structure, and Abstraction"
Science, 331(1279-1285), 2011, is grossly wrong, although I respect J. Tenenbaum.
Sorry, Josh.]

"These results show that single neurons can explicitly represent narrow, high-level,
natural stimulus categories"

[Weng: According to our computational model, this is incorrect, since such neurons
responding to "Saddam Hussein’’ are always many many. This shows again
that experimental biology alone is not sufficient to understand the brain.
Sorry, I respect experimental biologists.]

"The nature of the relationship between brain activity and mental representations
is a fundamental question in neuroscience, with relevance to disciplines ranging from
philosophy to cognitive science. While the answer in general is distant,"

[Weng: This is incorrect, not distant, but here and now. Collectively, the human
race seems to have sufficient knowledge to understand the relationship
between brain activity and mental representations. The major problem is isolation
of disciplines. Experimental biologists and experimental neuroscientists do not
have sufficient training in EE/AI/Math; but EE/AI/Math researchers do not have
sufficient training in Bio/NS/Psy. The planned Brain Mind Institute at MSU, whose speaker
series webcasts will start from this Fall, will serve everybody interested in this question.]

-John


On 6/23/11 10:06 AM, Asim Roy wrote:
Hi Peter,

Christof Koch forwarded your article titled “Neural Coding: Non-Local but Explicit and Conceptual.” The issue is grandmother cells and local representation. There are comments on your article by John Weng below. So it’s fair that we give you a chance to respond.

Here’s my question: As we parse this fact of selective response of single neurons and try to put it into a theoretical context, can we simply say, in the broader sense, that some single neurons carry substantive information, that they are not necessarily at the sub-cognitive level. I believe that’s the basic issue here when we talk about distributed vs. local representation. Local representation simply implies that neurons can encode information at the cognitive level. Let me know if that kind of broad generalization can be made from these single cell recordings.

Best,
Asim Roy
Arizona State University
www.lifeboat.com/bios/ex.Asim.Roy



On 6/23/11 10:23 AM, Walter J Freeman wrote:
It appears to me that the Jennifer Aniston cell is not so much a classical grandmother cell as it is a member of a Hebbian cell assembly, which performs inductive generalization over equivalent examples acquired in reinforcement learning. The finding demonstrates neither localization of a memory nor specificity of cell information, because it is likely that single cells can be members of more than one cell assembly, as one basis for association. What it does demonstrate is instant abstraction from an example to a concept.

The challenging question is how ~10^5 neurons capture ~10^10 neurons in a fraction of a second. A part of the answer comes from use of control system theory, which has shown that Hebbian synapses (positive feedback by mutual excitation) greatly facilitate the amplitude of mesoscopic gamma bursts (negative feedback by inhibition), which can broadcast the concept.

Best,

--
Walter J Freeman MD, Professor of the Graduate School
Department of Molecular & Cell Biology
Division of Neurobiology, Donner 101
University of California at Berkeley
Berkeley CA 94720-3206 USA
Tel 510-642-4220 Fax 510-642-4146
dfreeman@berkeley.edu
http://sulcus.berkeley.edu



On 6/23/11 10:54 AM, Jay McClelland wrote:

In a note by David Plaut and me, we made a few points about localist representation, in response to an attempt by Bowers to revive the Grandmother Hypothesis:

http://psychology.stanford.edu/~jlm/papers/PlautMcC10PsychRevReplyToBowers.pdf

Here are some of the key points and some comments on points from this discussion.

1. Intelligence would be too narrow if we had to rely on a neuron or set of dedicated neurons as the repository of the knowledge that we draw on when identifying or otherwise responding to items we encounter in our experience. Since many things are novel, we need a more general solution; and, we borrow information from other things that we know about to help us with gaps in what we know about specific things.

2. Even a system that appears to be localist in the sense that it contains neurons that respond more to one item than any other item may not be responsible for all the knowledge we draw on when we encounter that item, since other neurons activated to a lesser degree are very likely to still participate in influencing outputs (and generally that's a good thing, addressing the problem stated above, at least in part). The idea that the basis for the response I make to a given item arises solely through the consequences of activation of the dedicated population of neurons for that item is one that seems implicit in many discussions of the idea of grandmother cells, and was made explicit by Bowers in his commentary. Interestingly, however, the Interactive Activation model of letter perception (McClelland & Rumelhart, 1981), which is often cited as an example par excellence of a localist/grandmother cell model (and was so cited by Bowers), is not localist in this sense. Outputs in that model depended on the participation of many units -- when we process a string of letters, the unit for that word (if there is one), and the units for other similar words, all participate in the network's output. This is way the model addressed perception of letters both in familiar words and in novel but wordlike nonwords. To me what is most important is the notion that output depends on many units, no matter whether each unit corresponds to something 'explicit' (to use Foldiak's word) or not.

3. Localist theorists are exceedingly slippery when it comes to the definition of 'a thing'. Bowers' speaks of 'faces, words, and objects'. It is unclear what he means when he speaks of 'objects' -- are they particular objects (the Sydney Opera House) or classes of objects -- and if the latter, at what granularity? Christof's work (and his statement in the Scientific American) seems to suggest a privileged status for particular (familiar) persons, places, or things -- not to classes of things. Why the specificity should be at this level is interesting -- although it may be an artifact of the design of experment, or of the mentalities of the participants. Would they have found neurons that responded to 'german shephard', or 'scarlet taninger' or 'himmalayan snow tiger'? What if they had looked for such neurons in zoologists?

4. Learned distributed representations (and response properties of units within them) can vary in specificity, sparsity and explicitness (Foldiak's useful terms), sometimes approximating some of the characteristics of localist neurons. To Dave Plaut and me these are matters of degree, with localist or grandmother neuron sometimes serving as a useful approximation. Perhaps the neuron Christof et al found that responds to the Sydney Opera house and to another similar-shaped building points the the notion that indeed the grandmother cell characterization is only an approximation.

5. In one of Christof's papers the authors calculated that, given the number of neurons in the hippocampus and the number of items that we are familiar with, together with the specificity they observed, the average neuron they considered would be activated by 150 different items. The quote from the Scientific American seems to indicate that Christof now thinks otherwise. Or would Christof accept the following revision of the last part of the statement from the Scientific American?

The networks in the medial temporal lobe_recognize such repeating patterns and tune synaptic connections so that the neurons there respond more selectively to them_._Such selectivity applies to many types of repeating patterns, be they family members, pets, friends, co-workers, the politicians you watch on TV, your laptop, that painting you adore_. One and the same neuron, though, can become part of a population that is selective for more than one such pattern -- perhaps as many as 150 of them, in fact!


On 6/23/11 3:05 PM, Juyang Weng wrote:
Dear Jay,

Thank you for the points in your email. I agree with them, in general. However, I think that they are not powerful enough to explain why grandmother cells like those in symbolic Bayesian networks (e.g., in GOFAI and in Tenenbaum et al. Science 2011) are wrong, absolutely impossible for the brain-mind.

People do not use connectionist representation because people have not read the key literature or have read but do not understand (e.g., because of knowledge background) and therefore are still not sufficiently convinced. Let me play the devil's role below. I hope that this serious of emails is useful for discussion. To benefit more researchers, I will make all the related emails originated by Asim Roy under this Subject available at
http://www.cse.msu.edu/ei/IJCNN11panel/
This is the web site for an IJCNN 2011 penal that Asim and I are co-organizing.
In an accepted and upcoming IEEE TAMD paper "Symbolic Models and Emergent Models: A Review", I proposed to use "emergent representation" which seems more strict than the term "connectionist representation". The following ideas are discussed in that paper.

> 1. Intelligence would be too narrow if we had to rely on a neuron or set of dedicated neurons as the repository of the knowledge that we draw on when identifying or otherwise responding to items we encounter in our experience. Since many things are novel, we need a more general solution; and, we borrow information from other things that we know about to help us with gaps in what we know about specific things.

[Devil]: The same can be said for symbolic networks. They borrow information from other things through probability connections with other symbolic things.
[Weng]: Many things are novel, going beyond a finite number of combinations of a static set of hand-selected symbols. Symbols are for communication among humans based on human consensus, although each symbol (e.g., "cat") means different experiences in different brains. Therefore, symbols are outside the brain (not inside). Although the number of muscles that the brain can control is finite in life, the number of firing movies of muscles is astronomical, not bounded. For example, "Internet" is a new firing movie that the human society invented recently. This is the power of emergent representation (firing movies). It does not need the programmer to statically define a symbol for each new firing movie. In contrast, a symbolic approach requires a symbol for each firing movie. It does not allow any firing movies that do not have a corresponding symbol yet.

> 2. Even a system that appears to be localist in the sense that it contains neurons that respond more to one item than any other item may not be responsible for all the knowledge we draw on when we encounter that item, since other neurons activated to a lesser degree are very likely to still participate in influencing outputs (and generally that's a good thing, addressing the problem stated above, at least in part).

[Devil]: This idea is also in symbolic models. The ideas of pulling from a committee, where each committee member represents a symbolic concept.
[Weng]: A neuron cannot represent any symbolic concept. As the skull is closed during development, there is no way for a human teacher to assign a neuron
(or a specific set of neurons) to a symbolic concept that exists only in a consensual sense outside the brain, via communications among humans.

-John


On 6/23/11 3:12 PM, Asim Roy wrote:
Here's what Koch said in his article:

"One hippocampal neuron responded only to photos of actress Jennifer Aniston but not to pictures of other blonde women or actresses; moreover, the cell fired in response to seven very different pictures of Jennifer Aniston. We found cells that responded to images of Mother Teresa, to cute little animals and to the Pythagorean theorem, a2 + b2 = c2."

These are actual findings of several experiments I believe. Of particular note is the fact that these particular neurons encoded information about particular things. A single neuron was good enough to recognize an object (Jenifer Aniston, Mother Teresa) or a class of objects (cute little animals). There was no need to monitor or record the output of 150 other neurons to recognize Jennifer Aniston. In other words, there was no need to observe a pattern of activity across a cross-section of neurons to recognize Jennifer Aniston, even though it might exist in the system at various levels. I believe that's the crux of these findings. In general, it's fair to conclude from these experiments that single neurons can encode significant information at the cognitive level (e.g. recognizing Jennifer Aniston).

Asim



On 6/23/11 3:56 PM, Juyang Weng wrote:
> One hippocampal neuron responded only to photos of actress Jennifer Aniston but not to pictures of other blonde women or actresses

Although more detailed in hippocampus, this is similar to the V1 cells that Hubel and Wiesel reported, at least in principle. Do V1 cells recognize oriented edges? I do not think so, since they cannot recognize oriented edges, as many other cells also respond to similar oriented edges. According to my brain-mind model in Weng IJCNN 2010, each cell is a sample in a part of the joint space of X and Z, where X is the sensory space (e.g., retina) and Z is the motor space (e.g., muscles). This is true for all brain cells --- the spinal cord, the hindbrain, the midbrain, and the forebrain. Thus, it is true for V1 neurons, MTL neurons, and hippocampal neurons.
This is even true for different neurons in different laminar layers in the cortex, regardless it is inhibitory or excitatory. Each neuron does not represent
any meaning that can be explained precisely in human language.

> A single neuron was good enough to recognize an object (Jenifer Aniston, Mother Teresa) or a class of objects (cute little animals).

This is not true. For recognition of a class in the extra-body environment by a neuron (or it represents such a class), it must
(1) report whenever a sample in the class is present in the extra-body environment. Any single neuron cannot guarantee that.
(2) does not report when no sample in the class is present in the extra-body environment. Any single neuron cannot guarantee that.
Thus, no neuron recognizes any class (e.g., car) in he extra-body environment that can be expressed in a human language.

Computer science related courses (Automata theory, Pattern Recognition, AI, etc.) are useful for EE/Math/Bio/NS/Psy people, although they are not sufficient for understanding the brain-mind.

-John


On 6/23/11 9:24 PM, Ali Minai wrote:
Asim

Just because a neuron is tuned specifically to Jennifer Anniston and not to other blonde women does not mean that it may not also be tuned to a pumpkin (no disrespect to Jennifer Anniston). The firing of this neuron does not guarantee that Jennifer Anniston is being seen. Actually, showing that the neuron is tuned to Jennifer Anniston and only to Jennifer Anniston is really impossible for the same reason that proving the correctness of inductive learning is impossible. However, the experiments do show that specific neurons can be very narrowly tuned to specific stimuli, which tells us a lot about the nature of neural coding.

On your point "it's fair to conclude from these experiments that single neurons can encode significant information at the cognitive level", I don't think anyone would disagree. Hundreds of experiments over four decades have demonstrated this fact - from place cells and grid cells in the hippocampus to feature detectors in the visual cortex and cells tuned to specific objects or parts of objects in the temporal lobe. The maximal case of distributed representations that you're criticizing is a straw man - or at best a theoretical construct.

The more interesting issue is: What does the existence of specifically tuned neurons say about neural representations. Jay and John have both made excellent points on this. One intriguing piece of evidence is that whenever we go looking for a specific kind of tuning in the brain, we tend to find it. For example, neurons in the motor cortex are found to be tuned to direction of movement, velocity, muscle forces, postures, tasks, sequential position, etc., engendering much spirited debate among researchers. I think the obvious conclusion (which has been suggested by many) is that the system is not specifically encoding each of these things, but is encoding something more comprehensive of which these are subspace projections, revealed differentially by the nature of the experimenter's probe. At this point, we are like the five blind men trying to learn the nature of the elephant. We have separately discovered its various "natures", but are only beginning to put together the much more profound emergent truth that they represent in combination.

One thing that often gets lost in theoretical discussions is the physicality of the brain. We talk much of "embodiment" now, but usually apply that only to the body. Well, the brain is also part of the body, and is just as much subject to embodiment as any other part of the body. It has a structure that constrains its dynamics and endows it with emergent properties just as occurs in the musculoskeletal system. The brain's embodiment has far more deegrees of freedom, giving the illusion of infinite flexibility, but that isn't the case. Sometimes, as theorists, we sit around and discuss possibilities as if the brain could have chosen any of a set on the list using some objective fubction. In fact, this occurs through the optimization processes of evolution, development and learning, which are much more "on line", path-dependent and time-scale constrained than our disembodied theoretical studies. Evolution is a great optimizer, but a slow and messy one. Believing in evolutionary "rationality" is every bit as misleading as believing in economic "rationality". We should consider the physical system before us and investigate its emergent properties to understand what it does and how.

All Best

Ali



On 6/24/11 1:50 AM, Asim Roy wrote:
The simple fact is that single MTL neurons can encode information about an object in a comprehensive way. (Koch: “One hippocampal neuron responded only to photos of actress Jennifer Aniston but not to pictures of other blonde women or actresses; moreover, the cell fired in response to seven very different pictures of Jennifer Aniston.”). There was no need to read the outputs of 150 other neurons and interpret them according to some encoding scheme to recognize these objects and concepts. Moreover, the experiments have been repeated for several different objects and concepts (Koch: “We found cells that responded to images of Mother Teresa, to cute little animals and to the Pythagorean theorem, a2 + b2 = c2.”) . These are not isolated or rare findings. Single cell recording experiments have been going on for some years in neuroscience and Koch’s article in Scientific American Mind summarized some of those findings. And the fact that the firing of these particular neurons can be associated with particular objects or concepts indicate that these neurons are at the cognitive level, not at the subcognitive or subsymbolic level.

Again, these findings clearly show that we don’t need to read out a pattern of activity across hundreds of neurons (150 by some estimate??) to recognize objects and concepts. There are single neurons that can perform that function. And these experiments are repeatable.

Asim



On 6/24/11 9:39 AM, Asim Roy wrote:
Dear Walter,

You know cell biology better than any of us. Given your interpretation, how would you define a “concept?” Is “Jennifer Aniston” a concept? And where does the “concept” exist and in what form? This question relates to the “instant abstraction” that you mention.

Best,
Asim


On 6/24/11 9:59 AM, Juyang Weng wrote:
> Koch: “One hippocampal neuron responded only to photos of actress Jennifer Aniston but not to pictures of other blonde women or actresses; moreover, the cell fired in response to seven very different pictures of Jennifer Aniston.”)

This finding and many other findings of similar nature are consistent with my brain model: Every hippocampal neuron, like any other neurons in the Central Nervous System, has a weight vector as a cluster in part of the joint space of X and Z, where X is the sensory space (e.g., retina) and Z is the motor space (muscle array). It is the "part" term that makes the brain very complex. A major reason why Koch's particular neuron responded the " Jennifer Aniston" is the Z component of the neuron, where Z drives, e.g., the vocal tract contractions that produce the sound "Jennifer Aniston". This mechanism is explained in Miyan & Weng ICDL 2010.

> There was no need to read the outputs of 150 other neurons and interpret them

This is not true, since the brain must produce required action as a result of recognition. In the above example, the brain must pronounce "Jennifer Aniston". Many neurons respond to some "Jennifer Aniston" specific aspects but also respond to "Jennifer Aniston" nonspecific cases. Koch did not check whether that particular neuron also responds to some "Jennifer Aniston" nonspecific cases, but my theory predicts that it must do! How does the brain generates the action that reports "Jennifer Aniston" cases? This is an open problem for the brain. My model in IJCNN 2010 explains how the brain does it. My three theorems to be presented in IJCNN 2011 in San Jose proves that my model (Developmental Network DN) can be error-free for training data and optimal for testing data.

> these neurons are at the cognitive level, not at the subcognitive or subsymbolic level.

These individual neurons by themselves cannot solve cognition. Generate such so-called "cognitive level" neurons are not difficult. Simply randomly sample the joint space of X and Z. You will get plenty of them!

Experimental biology/neuroscience has made advances due to availability of new technology (e.g., single cell recording). However, these two disciplines alone cannot answer key brain-mind questions. Our 5-chunk model with experimental verification was submitted to Nature during the same period as Koch's paper. My model not only computationally explains Koch's finding but also explains how the brain works to give general-purpose solutions, although I did not know about Koch's paper when we submitted our paper. I applaud Christof. Our paper to Nature was rejected.

After I have presented my 5-chunk brain-mind model in IJCNN 2010, it will take some time for different disciplines to digest, understand, and accept my model. The current situation is that each individual researcher in a related discipline (e.g., Bio/NS/Psy/CS/EE/Math) does not have sufficient background knowledge to understand my model or any other correct brain model in the future. Do not worry though, as MSU's administrators seen this problem and is willing to help us with the Brain-Mind Institute from this Fall.

-John


On 6/24/11 5:59 PM, jeff bowers wrote:
Thanks Asim for including me in this exchange. I expect my interest in the grandmother cell hypothesis is motivated someone differently from most people here. I’m a cognitive psychologist, and I have been struck how many people prefer PDP models over “localist” ones (e.g, the IA model of McClelland) because PDP models are considered more biologically plausible. The key feature of localist models is that single units represent one category (e.g., the word DOG) and as a consequence, responds to high level information in a highly selective way (e.g., a DOG unit responds highly selectively to the word DOG). What I have been trying to highlight is that the selectivity of localist representations is consistent with what is found in single-cell neurophysiology studies. PDP models, by contrast, have traditionally assumed that very little information can be had b examining single hidden units, and as a consequence, there is very little work reporting single-unit recordings in PDP models. On the assumption that PDP models are more biologically plausible, it would be expected that they would capture the data better (in fact localist models do a much better job capturing the single-cell recoding findings).
It is interesting to me that prominent computational neuroscientist develop localist models – e.g., see: Maximilian Riesenhuber and Tomaso Poggio in Nature Neuroscience (2000) where they write:
“We propose a model (Fig. 3) that extends several existing models 5,39,40,42,43. A view-based module, whose final stage consists of units tuned to specific views of specific objects, takes care of the invariance to image-based transformations. This module, of which HMAX 42 is a specific example, probably comprises neurons from primary visual cortex (V1) up to, for instance, posterior IT (PIT). At higher stages such as in anterior IT, invariance to object-based transformations, such as rotation in depth, illumination and so forth, is achieved by pooling together the appropriate view-tuned cells for each object.”
They in other places take their model as an extension of the simple-complex cell hierarchy proposed by Hubel and Weisel – just the sort of hierarchy included in localist models (but not common to PDP models). I should also say that Riesenhuber and Poggio do not consider their theory a grandmother cell theory – but that is only because they have developed a very extreme definition of a grandmother cell theory. That is, according to these authors, a grandmother cell responds to one and only one category of object (e.g., a specific face) – with no activation to anything else. I don’t see why such an extreme view needs to be taken – it is not the way localist units work (the localist unit for the word DOG is highly selective to the word DOG, but it will fire somewhat to form related words like HOG, FOG, etc.) But in any case, they have developed a model that is very much like the IA model of word identification, but applied to faces (with a single unit tuned to a specific face). It makes it difficult to argue that a PDP model of word identification is more biologically plausible that a localist one.

Some of the comments above point out that even though some neurons have been found that responded to only one image out of 100 or more images sampled, the neuron undoubtedly will fire to other objects – estimates of about 150 other objects have been suggested. This is taken to be problematic for localist (or grandmother) cell theories. But in fact, most (if not all) localist models predict that a given unit codes for one thing, but will be activated by other things as well. In case anyone is interested, here is a recent paper of mine that attempts to address some of what I see are confusions on this topic (including confusion in a paper by Foldiak that was mentioned earlier).

Bowers (2011) What is a grandmother cell? And how would you know if you found one? Connection Science, Volume 23, Issue 2, 2011, Pages 91 - 95
http://www.informaworld.com/smpp/content~db=all~content=a938092195
Jeff


On 6/24/11 8:46 PM, Juyang Weng wrote:
Jeff,

I enjoyed reading your views. PDP models and "localist" models all belong to the class of connectionist models, I think.

Although commonly used, the term "connectionist" is misleading and confusing. A symbolic network such as a Finite Automaton and its probabilistic variants (which I call Symbolic Nets --- SN --- such as HMM, POMDP, belief nets, Bayesian nets, semantic nets, and graphic models) all have a network. One may provide a long list to explain many possible differences between a connectionist (PDP) model and an SN. However, let us consider whether they are brain-like networks
using the following fact.

The brain is skull-closed after the birth. It does not allow the teacher/mother to open the skull and directly twist the internal representation (symbolic, localist, or distributed!). Any brain model, regardless of a connectionist model or an SN, must explain how this network emerges autonomously inside the skull of the brain throughout the life, learning one new concept after another from the environment (e.g, Internet which cannot be gene fully specified!), while the brain interacts
with the external world through its two ends that are exposed to the external world --- the sensory end (e.g., retina) and the motor end (e.g., muscle array).

I do not think that any SN can do that, because the handcrafted symbolic concepts. I think a connectionist model has a chance, but only if its internal representation
is fully emergent. However, the requirement of fully emergent internal representations raises a great challenge on network capabilities. Prior fully emergent networks (e.g., SOM) are limited in capabilities --- they cannot learn immediately (they iterate many times to converge), cannot be error-free on the training data (only approximate, but an adult brain can be immediately error free on a training case), and they cannot do state-based logic (but a brain and a Finite Automaton can).

Furthermore, if we require that the brain is skull-closed after birth, the only grandmother cells are at the motor end. Yes, in the strict sense --- a grandmother cell responds to one and only one category of object (e.g., in the motor area of our skull-closed Developmental Network DN). Your vocal tract responds correctly to one and only one category of object (e.g., human face). This is because the motor end is open to, and is observable by, the environment, so that the human society and the agent (e.g., a child) can work together to calibrate each action from the motor.

However, I argued that from a biological brain even a calibrated action is not exactly the same every time (e.g., the word "yes" from my vocal tract is at least slightly different each time (e.g., slightly different tones.) Therefore, the strict sense of a grandmother cell does not seems to exist in the skull-closed brain at all.

If one does not talk about grandmother cell in the strict sense, then the loose concept of grandmother cells is not very useful, as such cells are everywhere in the brain.

The above views are raised and discussed in my accepted and upcoming IEEE-TAMD paper "Symbolic Models and Emergent Models: A Review".

-John



On 6/25/11 12:43 AM, Ravi Rao wrote:
John,

You've repeatedly brought up the issue of isolation of disciplines.

"The major problem is isolation
of disciplines. Experimental biologists and experimental neuroscientists do not
have sufficient training in EE/AI/Math; but EE/AI/Math researchers do not have
sufficient training in Bio/NS/Psy." (from your note attached).

I absolutely agree with you. This issue goes beyond the basic training that you've brought up.
There also needs to continuous interaction between the researchers in these disciplines, which
is hard to achieve. For instance, I rarely see attendees of IJCNN
going to the Society for Neuroscience Annual Meetings or vice-versa.

There are efforts to promote interdisciplinary interaction, and amongst them is
the National Institute for Mathematical and Biological Synthesis.
They provide opportunities for focused interactions, and I'm mentioning this
as it may be of benefit to others on the discussion list.
You may visit http://www.nimbios.org.
I created a group to investigate cortical networks: http://www.nimbios.org/workinggroups/WG_CorticalNetworks
I'll be at the IJCNN meeting and would be happy to discuss this with anyone who is interested.

Best regards,

Ravi


On 6/25/11 3:01 AM, Asim Roy wrote:
Hi Jeff,

Thanks for the note. I think the most important finding in these single-cell recording studies is that the firing of particular neurons can be associated with particular objects or concepts. That indicates that the output of a single neuron can have meaning (semantics) and thus they are at the cognitive level. A very fundamental claim of connectionism is that such neurons couldn’t exist, that all neurons and processes are at the subcognitive or subsymbolic level. So the existence of such neurons is very revealing.

Second, these findings clearly show that we don’t (or the brain doesn’t) need to read out a pattern of activity across hundreds of neurons to recognize objects and concepts. There exists single neurons that do that read out internally and captures the result of all of the underlying neuronal activity. Thus there is consolidation of information in a single neuron and that too is revealing.

I think these findings are really important for us.

Asim



On 6/25/11 4:50 AM, Ali Minai wrote:
Asim

I don't think that the first item you cite is "a fundamental claim of connectionism", though it has been made on occasion by some. All in the field (including most connectionists) have known for a long time that cells are tuned to specific stimuli, location, concepts, features, etc. Why are we still talking about this? The interesting thing in the new results in the degree of specificity, and we still have a lot to learn on how information is actually encoded in the brain.

And the findings do not "clearly show that ... the brain doesn't ... need to read out a pattern of activity across hundreds of neurons to recognize objects or concepts." For one, "the brain" does not "recognize" anything any more than feet walk home or hands write letters. These are the actions of the animal/human. The brain, like the rest of the body, just undergoes its natural dynamics, in the process of which emerge phenomena that we, as observers, call "recognition", "thought", "memory", "decision", etc. The pertinent question, therefore is not what the brain needs to read out, but what we would need to read out to infer that the person whose brain we are monitoring is seeing Jennifer Anniston. Very concrete cases of this are currently being studied intensively by people working on brain-machine interfaces, who are asking questions like:

How many cells in the motor cortex need to be read from to infer what movement is to be generated by the hand?
How many (and what) cells in the cortex need to be monitored to determine what object a person is thinking of?
How many hippocampal place cells must be monitored to determine where a rat thinks it is?

These questions are being answered very concretely by people like Andy Schwartz, Eric Leuthardt, Ted Berger, Jose Carmena and many others. Jose and Ted will be speaking at IJCNN, so that would be a good opportunity to discuss this.

If this group is to engage is useful discussion, I think it is very important to move beyond imagined "straw man" positions that no one actually holds. The idea that individual neurons do not encode substantially specific information is one of these imaginary positions. Almost no one believes it, so why are we arguing about it? And to say that the only viable alternative to this is total localism is a fallacy. There is a vast range of positions between total distributedness and total localism, and that is where the truth lies.

Best

Ali



On 6/25/11 5:28 AM, Asim Roy wrote:
Hi Ali,

I hope you are right, that it’s a “straw man” position. I am sure many connectionists would disagree with you. There are very recent papers on these questions starting with the article by Bowers in Psychological Review. These papers will give a feel why Koch’s findings are troubling to many connectionists.

Asim



On 6/25/11 7:35 AM, Walter J Freeman wrote:
Dear Asim,

Thank you for your questions. I use the dictionary definition: 'concept' is an abstract idea. "Jennifer Aniston" is not a concept. It is a phrase symbolizing the idea of the person.

From philosophy a concept is "an idea or mental picture of a group or class of objects formed by combining all their aspects". The aspects preceding a concept are unimodal percepts formed from sensations. In evolution the prototypic percept is olfactory. In fuzzy numbers there are 10^8 receptor cells with 10^3 types, hence 10^5 of each type. Each sniff excites 10^2 cells. The cells sampled differ on every sniff. How does the olfactory bulb generalize over equivalent cells? Lashley first identified the problem, calling it "universal". Hebb solved the problem, using the connectionism of Cajal and Lorente de Nó.

Where and what form? The stereotypic output of the Hebbian assembly triggered by each sensation is distributed in the entire assembly of 10^5 neurons wherever they are. An assembly is unique to each subject. It combines experiences stored in the synapses. The output is an abstraction because the sensory detail is expunged on convergence to the attractor. It is the first step in the making of mind.

In my opinion, this is as far as connectionism can go. The next step is a phase transition from a sparse mesoscopic representation to a dense macroscopic representation, resembling a condensation. Physics takes over from network theory. I'll talk about that in my tutorial in San Jose.

Best,
Walter



On 6/25/11 9:52 AM, Asim Roy wrote:
Dear Walter,

You are in the best position to reconcile these findings from single cell recordings (Koch: ““One hippocampal neuron responded only to photos of actress Jennifer Aniston but not to pictures of other blonde women or actresses; moreover, the cell fired in response to seven very different pictures of Jennifer Aniston. We found cells that responded to images of Mother Teresa, to cute little animals and to the Pythagorean theorem, a2 + b2 = c2.”) with the attractor theory. It’s all at the cell level, what you might term the mesoscopic level.

Best,
Asim

On 6/25/11 11:06 AM, Juyang Weng wrote:
Dear Water,

Your views are probably well accepted by many biologists and neuroscientists.

> The output is an abstraction

Well, I think that it is partially true for, e.g., "type" motor output (e.g., via vocal tract) such as saying "car". It is not true for "manipulatory" motor outputs such as dancing and navigation. In fact, the vocal tract output is also manipulatory --- firing patterns of motor neurons in the vocal tract. Our DN model take into account all types of motor outputs that depend on sensorimotor contexts (not just sensory context), including "type", "location", spatiotemporal events, and all other possible concepts.

> this is as far as connectionism can go

In our Where-What Networks (experimental examples of DN), connectionism (must be fully emergent from the closed skull) goes everywhere in the brain, to do abstraction, reasoning, goal directed search, language acquisition and understanding, etc.

> The next step is a phase transition from a sparse mesoscopic representation to a dense macroscopic representation, resembling a condensation.

This is a stage view about the brain, limited in biological causality. Our DN model in Weng IJCNN 2011 does not think that the brain has rigid step 1, step 2, etc. I used such stepwise representations till around 2005. After that, theoretical conceptualization from rich neuroscience fasts helped me to see the deeper causality: The detailed connection diagrams such as the large and complex ones put together by Felleman & Van Essen 1991 from over 50 studies are misleading. The brain does not have a static partition of areas or a rigid connection diagram. The diagram that Felleman &Van Essen 1991 put together seems the major statistics discovered by the animals, not all the statistics but major ones. Weng IJCNN 2010 explained these views that motivated our 5-chunk brain-mind network model with citations.

D. J. Felleman and D. C. Van Essen (1991), "Distributed hierarchical processing in the primate cerebral cortex", Cerebral Cortex, vol. 1, 1-47, 1991.

-John



On 6/25/11 1:57 PM, Jay McClelland wrote:

I've been trying to keep my contribution to this discussion to a single message, but the recent exchange has (for better or worse) led me to re-engage.

First, I'd like to applaud Ali's points -- in my view, they are right on target.

I also wanted to remind Asim that Chistof's estimate is that the Jennifer Anniston neuron responds to Jennifer and about 150 other things. If this is true, your assertion that his findings show that the brain doesn't need to read out a pattern of activity across hundreds of neurons might not be correct. The response properties of these neurons are very similar to the response properties that neurons would have in the kind of sparse, random conjunctive representations David Marr proposed might be used in the hippocampus (many others, including me, Bruce McNaughton, and Randy O'Reilly have since argued for this kind of representation as well). The data provide some challenges to existing models of this type and I would love to see someone work on them further to address all aspects of Christof's (and other people's) data on single neuron response properties in the medial temporal lobe.

Further, it is important to note that humans can recognize and access knowledge about individuals, words, ideas, theorems, and object categories without having a medial temporal lobe. The medial temporal lobe is only necessary for new learning. Other parts of the brain (which, if our hypotheses are correct, use denser distributed representations) subserve everyday conceptual knowledge in its everyday form. In any case, the important feature of distributed models is their emphasis on the similarity relations among patterns of activation over units, not on the units themselves. A paper by Kiani et al (cited below) is consistent with the view that patterns over real neurons in areas outside the MTL have the kind of similarity relationships we see in learned distributed representations. It would be interesting to have data on large populations of neurons in human MTL, instead of having to rely on a few isolated neurons.

A small terminological point: The term connectionist models encompasses localist and distributed models. The term was introduced by Jerry Feldman and Dana Ballard, and Jerry was (and still is. I think) a champion of localist models. The phrase "parallel distributed processing" was coined by Rumelhart and me to distinguish a particular class of connectionist models -- those that use distributed representations -- from the localist kind. It certainly is not a fundemental tenet of connectionist modeling that representation must be distributed. I am of the view that localist models are often useful as approximate characterizations but when one looks more closely at certain details, they cease to be sufficient, in just the way Newtonian physics often provides a useful approximate characterization in physics. For reasons stated above, I don't think Christof's findings actually challenge the tenet that mental representatations (even in the MTL) are distributed patterns of activation over populations of neurons.

Although Bowers adopts the stance that the justification for PDP models is based on the claim that they are more biologically realistic than distributed models, I would disagree. The primary appeal for me is the range of new ideas they offer for understanding how experience shapes cognition and behavior, and how performance in cognitive tasks degrades with damage to neocortical brain areas. With all do respect to localist models, the localist models generally don't learn -- they are wired to have certain knowledge in them by a programmer. Thanks to Rumelhart, Hinton, and others, we actually have models now that learn powerful distributed representations that can subserve cognitive performance like that of humans in many tasks, and simulate patterns of breakdown of human performance after brain damage. And, as stated above, the models learn to assign related patterns to related concepts in ways that capture the similarity relations among the patterns of activity used in the brain.

Jay McClelland
On 6/25/11 3:55 PM, Asim Roy wrote:
Here’s the process of formation of what Koch calls “concept neurons”:

“Every time you encounter a particular person or object, a similar pattern of spiking neurons is generated in higher-order cortical regions. The networks in the medial temporal lobe recognize such repeating patterns and dedicate specific neurons to them. You have concept neurons that encode family members, pets, friends, co-workers, the politicians you watch on TV, your laptop, that painting you adore.”

If this process, or at least the result of the process, has been experimentally verified, then it appears that the brain is recognizing repeating patterns and creating separate neurons to recognize each such repeating pattern. So it appears that the read out of repeating patterns is internalized within these dedicated neurons. Common sense tells me that’s a smart and efficient way to operate; in the long run, it minimizes the effort required to interpret patterns that repeat. And the read out of repeating patterns is very much simplified – it’s flagged by the output of a few dedicated neurons. That’s what the experimental findings show. And I guess dedicating neurons to repeating patterns is part of the learning, they are not hard-wired.

I am copying this to Dr. Itzhak Fried at UCLA, where much of the experimental work was done. I am sure he and Christof Koch can better clarify some of these interpretations of their findings.

Asim Roy



On 6/25/11 4:24 PM, Giacomo Indiveri wrote:
Dear Asim,

Here's a comment from a naive observer of the discussion that has been
going on, without any particular bias: I think the argument arises from
the fact that you confuse the observation that neurons "respond to"
concepts of Jennifer Anniston, with the conclusion that they "recognize"
Jennifer Anniston.

As many have been implying in their replies, the two things are quite
different.

You start your replies by stating clear unambiguous facts (such as the
results of the experiments), that everybody agrees on. But then you
conclude your emails making interpretations that are quite arguable (and
are being argued upon), saying that the facts "clearly demonstrate" that
single neurons can recognize objects, concepts or whatnot.

Probably the best course of action at this point is to take this
discussion off-line, collect all of the statements and arguments that
have been made, and use them in the panel that you will hold at IJCNN.
I am sure it will be a really lively and interesting debate!

Best regards,

giacomo




On 6/25/11 4:33 PM, Asim Roy wrote:
Dear Giacomo,

Thanks for the suggestion. This can go on and on and may not be of interest to all on this list. Will take it offline.

Best,
Asim

On 6/26/11 4:56 AM, Walter J Freeman wrote:
Š and let the "grandmother cell" retain its intended and historic connotation of absurdity.

Walter


On 6/26/11 5:01 AM, Walter J Freeman wrote:
Dear John,

I think we are in basic agreement. We should explore this at leisure some evening, perhaps at the IJCNN.

Best,

Walter


On 6/26/11 7:18 AM, Jeff Bowers wrote:
Sorry if this is one too many comment, but I can’t resist responding to a few points above.

Walter Freeman writes: “ let the "grandmother cell" retain its intended and historic connotation of absurdity.”

No doubt the term “grandmother cell” is pejorative, and indeed, some versions of this hypothesis are absurd (the idea that one and only one neuron fires in response to a particular face, a particular word, etc). But in my view, some versions this hypothesis are from absurd, distinct from “distributed” theories, and should not be dismissed so quickly. For example:

Should Poggio and colleagues theory of face identification that is functionally similar to the IA model (with single units tuned to specific faces) be dismissed as absurd? If not, should the IA model of word identification be considered less biologically plausible that various PDP models of word identification?

Should Grossberg’s ART models of categorization, and winner-take-all networks more generally, be considered absurd (a model that learns its categories?)
These are the types of models that I think should be taken seriously, both as cognitive models, and as biological models. But the dismissal of “grandmother cells” is often used as an argument against these types of theories.

Ali Minai writes:

“All in the field (including most connectionists) have known for a long time that cells are tuned to specific stimuli, location, concepts, features, etc. Why are we still talking about this? The interesting thing in the new results in the degree of specificity, and we still have a lot to learn on how information is actually encoded in the brain.”

The key question is whether a given cell is tuned to a specific stimulus (one as opposed to many different stimuli)i. If so, I think this is an argument for local (grandmother cell) representations. A key claim of distributed representations is that a cell is not tuned to one stimulus, but rather, is tuned to a collection of different (perhaps related) stimuli. Each neuron is involved in coding different stimuli, and it is the pattern that defines a given stimulus. In the Poggio model, a cell is tuned to one specific face – that is, it represents a given face. On my terminology, that makes it a localist (grandmother cell theory). I agree that many connectionists have endorsed localist representations. A key claim of PDP modelers (a specific version of connectionism) was that localist representations (as in the IA model, or in Grossberg’s models) should be rejected.
McClelland writes:

“Although Bowers adopts the stance that the justification for PDP models is based on the claim that they are more biologically realistic than distributed models, I would disagree. The primary appeal for me is the range of new ideas they offer for understanding how experience shapes cognition and behavior, and how performance in cognitive tasks degrades with damage to neocortical brain areas. With all do respect to localist models, the localist models generally don't learn -- they are wired to have certain knowledge in them by a programmer.”

However, many prominent PDP advocates do take their biological plausibility as a key feature (as important as their ability to account for behavioral data). For example, Seidenberg and Plaut (2006) , when contrasting “localist” and distributed models, note that localist models often provide a better account of the relevant data:

“The PDP approach to cognitive modelling in general, and word reading in particular, carries with it a number of implications that are worth spelling out in detail. The first and most obvious is that the development of a model is subject to several constraints, only one of which is fitting specific behavioural findings. The result is that, in the short run, specific PDP models may not match particular empirical findings or account for as much variance in empirical data as approaches for which data fitting is the primary goal. “

So why prefer PDP models? It is their neurobiological plausibility:

“The PDP approach is different. The models are only a means to an end. The goal is a theory that explains behaviour (such as reading) and its brain bases. The models are a tool for developing and exploring the implications of a set of hypotheses concerning the neural basis of cognitive processing. Models are judged not only with respect to their ability to account for robust findings in a particular domain but also with respect to considerations that extend well beyond any single domain. These include the extent to which the same underlying computational principles apply across domains, the extent to which these principles can unify phenomena previously thought to be governed by different principles, the ability of the models to explain how behaviour might arise from a neurophysiological substrate, and so on. The models (and the theories they imperfectly instantiate) aspire to what Chomsky termed “explanatory adequacy”. The deeper explanatory force derives from the fact that the architecture, learning, and processing mechanisms are independently motivated (as by facts about the brain) rather than introduced in response to particular phenomena.”

This is the claim I’m challenging. As for the claim that localit models do not generall learn – the more relevant point is that many localist (winner-take-all) networks do. So the question for both localist and distributed PDP modellers is how well their models account for the data,how well they learn, and how well they account for the neuroscience. Perhaps the brain is more like a PDP model than a localist model – but the dismissal of absurd versions of the grandmother cell theories provide little evidence that PDP models should be preferred (or distributed theories more generally) .

jeff


On 6/26/11 8:31 AM, Dick Windecker wrote:
Hi Ravi,

I’m with you 100% on that one: the isolation of disciplines. I hope we can find some time during the meeting to talk.

This is exactly the reason I am giving a tutorial on “stochastic artificial neurons and neural networks” at this year’s meeting. At IJCNN 2009 I became convinced that for the most part, neither the biologists nor the engineers and computer scientists understood as much as they should about the mathematical foundations of probability theory and of complex stochastic systems constructed from simple stochastic parts.

Of course, as a physicist, I have to admit that I am definitely a (tiny) part of the overall problem: I certainly don’t know as much as I should both about biology and about engineering and computer science. But I think I can at least contribute in a small way toward the solution. Hence the tutorial.

By the way, I am also giving a poster presentation on random walks. This is mostly to provide an example of how stochastic ANNs can be applied and, at the same time, possibly provide some insights on how living systems might work. But I also see random walks as indirectly related to autonomous learning in the sense that autonomous learning sometimes involves searching and searching sometimes involves random walks.

Anyway, one of the things I do like about IJCNN is that I get to talk to people from both the technical and the biological communities, and from both the experimental and the theoretical sides. This does lead to certain feelings of schizophrenia, but I also find it very stimulating. On the other hand, as you say, much more needs to be done to really address the problem of the isolation of disciplines. I have recently had direct experiences of this problem in talking to some of the experimental biologists studying random walks in various animals.

Best,

Dick Windecker



On 6/26/11 4:10 PM, Harry Erwin wrote:
(Second half of the addressee list. We really need a list server.)

In my research into bat behaviour, I encountered evidence for a number of interesting things. One was that the bats I was working with did not appear to be stimulus-response machines. It seemed more like they had a model of the world, updated asynchronously, and they used that to choose their behaviour. Another was that they could rapidly simulate non-linear trajectories in time--taking account of all nine dimensions defining the object state. This all takes place in a two-dimensional sheet of neurones, and topologically, there can be no one-to-one association between 2-dimensional neighbourhoods in the brain and 9-dimensional neighbourhoods in the outside world, so the brain has to use a distributed representation in that context. Understanding how that dimensionality reduction takes place might be a productive research topic.

Incidentally, more recently, I've been working with Gibbs sampling, a Bayesian modelling technique that works from conditional distributions rather than the full joint distribution. I've become intrigued by the idea that it might give insight into how the brain models reality. In particular, it can be used with an asynchronous input stream of data to maintain a estimate of the current state of the world. I'm wondering if it might also be adapted to explaining how trajectory prediction could be done in an implicit rather than explicit way.



On 6/27/11 1:51 AM, Asim Roy wrote:
Hi All,

Let's do this offline at this point. I think it has been an interesting exercise given Koch's recent article that should make us think. As one might see, there still are a variety of different positions on the representation issue. And the discussions brought it out clearly.

Please email me if you want to be included in any offline discussion on this issue.

Thanks,
Asim


On 6/27/11 2:09 AM, Asim Roy wrote:
Dear Walter,

"Concept cells" sound interesting. That's also the term used by Christof Koch in his article. He doesn't really use the term "grandmother cell." I was wondering if you might be interested in writing a short note for some journal defining the term "concept cell" and what it might mean. I think we sometimes argue over terms that are never properly defined anywhere. I am sure some of us would like to participate in defining a "concept cell," myself included. I can check with a few others.

Let me know if you think this might be a good exercise. This would also give us a chance to reconcile the various theories and models with the new findings in neuroscience.

Best,
Asim



On 6/27/11 2:26 AM, Peter Foldiak wrote:
Asim,
I think density of encoding (ranging from locality to fully distributed)
is a property measured across the neurons (either for single stimuli, or
averaged across stimuli), while breadth of tuning is a property is the
property of a neuron measured across (ranging from grandmother-like to
broad).
But even a dense code can have (some) grandmother-like cells (which
makes sense for high frequency stimuli), and a highly sparse code can
have (some) broadly tuned neurons.[1] There is no contradiction between
the two.

An interesting solution is to apply Formal Concept Analysis (FCA) to
neural codes.[2] The input to FCA is a matrix representing which neurons
responds to which stimuli. The output is a semantic graph (a lattice) of
the related concepts that clearly show which subsets of stimuli are
coded by which subsets of active neurons. Here, a "concept" is just a
pair of sets of objects and sets of active neurons. (This can be
generalised to graded responses too [3].)
I think FCA can replace the ideology in these discussions with more
practical answers.

[1] https://www.st-andrews.ac.uk/~pf2/FoldiakCurrentBiology2009.pdf
[2] http://www.st-andrews.ac.uk/~pf2/EndresFoldiakNIPS2008.pdf
[3] http://www.st-andrews.ac.uk/~pf2/AMAI-EndresFoldiakPriss.pdf

(sorry if too many people got copies of these emails)


On 6/28/11 7:26 PM, Juyang Weng wrote:
Dear Walter,

In addition, I do not think that there is a master map in the brain, e.g., p. 504 in Kandel, Schwartz and Jessell, Principles of Neuroscience,
originally proposed by Anne M. Treisman and used by Anderson & Van Essen 1987; Olshausen, Anderson & Van Essen 1993; Tsotsos 1995; and others.

I will enjoy discussing with you during IJCNN. I will be there.

Best regards,

-John



On 6/28/11 8:17 PM, Juyang Weng wrote:
Those who do not like to be included in this discussion please ask Asim to remove.

Walter and Aim,

I do not think that "Concept cells" is an appropriate term either. According to the brain anatomy and connection patterns in the literature (e.g., Felleman & Van Essen 1991), almost all the cells have three input parts, a bottom-up input part which is typically from receptors, a lateral input part which is typically from cells in the same area, and a top-down input part which is typically from the motor areas. Every neuron is a sample of these three spaces. Of course, a neuron cannot fire simply based on its three parts of input. Many neurons compete to fire via inhibitory connections. If it fires, its match on these three parts is among the top.

That is how my 5-chunk model in IJCNN 2010 mathematically modeled. It has been mathematically and experimentally shown to have surprisingly powerful functions when these neurons work together, such as immediate learning and error free in supervised learning when tested using train data (i.e., the brain-like network does not need to iterate slowly to converge, one example is sufficient). That paper stated that if the model is a good approximation of the natural brain, than almost no single neuron in the brain can represent any concept (or meaning) that corresponds to anything that can be precisely described by any human language. In order words, there seem no "Concept Cells" in the brain. Sorry, Christof, if you used the term.

Best,

-John


On 6/29/11 5:12 AM, Asim Roy wrote:
Hi John,

Let's do this offline. I will take a look at your 5-chunk model before San Jose. On concept cells, I think there is a growing body of evidence for it from neuroscience. Can't simply deny it. So we have to understand what it means for the various theories and models that exist.

Best,
Asim



On 6/29/11 8:52 AM, Walter J Freeman wrote:
Dear Walter,

In addition, I do not think that there is a master map in the brain, e.g., p. 504 in Kandel, Schwartz and Jessell, Principles of Neuroscience,
originally proposed by Anne M. Treisman and used by Anderson & Van Essen 1987; Olshausen, Anderson & Van Essen 1993; Tsotsos 1995; and others.

John
I agree, there is no 'master map'. Or master cell.

I think you missed my point that these are Hebbian cells - each is a member of a set, all of which fire when any fires, hence inductive generalization. Hence concept.

See you in San Jose,

Walter


On 6/29/11 10:18 AM, Juyang Weng wrote:
> Hebbian cells - each is a member of a set, all of which fire when any fires, hence inductive generalization. Hence concept.

According to the skull-closed requirement during brain development, this kind of hand-crafted logic-like relationship among cells do not seem to exist in the
brain either. The teacher mother cannot open the skull to wire cells in such wishful thinking logic. In fact, this kind of wishful thinking logic is not desirable, e.g.,
run-away complexity. Our DN model demonstrated general-purpose functions using only cell mechanisms such as STDP (Hebbian learning) and Synapse Maintenance (growth and retraction of synapse to get rid of background components in the receptive fields that are not related to the post-synaptic firing, to appear in Wang, Wu, and Weng IJCNN 2011).

-John


On 6/29/11 2:24 PM, Juyang Weng wrote:
Sorry. Could you explain a little more about

> Hebbian cells - each is a member of a set, all of which fire when any fires, hence inductive generalization. Hence concept. ?

Are you talking about "member neurons" and the "set" neuron, where the "set" represents all the members?

-John
On 6/30/11 8:56 AM, Walter J Freeman wrote:
Say your stimulus excites 100/100K member neurons, the whole set ignites. Doesn't matter which 100. A neurobiologist stumbles on one at a time. Any will do.

Hard part is to explain how those 100K take control of the entire cortex - see attached in press TBME

Walter


On 6/30/11 11:32 AM, Juyang Weng wrote:
Good point. Let us consider this scenario.

My stimulus (cat view 1) excites 100/100K member neurons, called set S1. The whole set S1 ignites.
My stimulus (cat view 2) excites 100/100K member neurons, called set S2. The whole set S2 ignites.
My stimulus (dog view 1) excites 100/100K member neurons, called set S3. The whole set S3 ignites.
My stimulus (dog view 2) excites 100/100K member neurons, called set S4. The whole set S4 ignites.
On and on ...
Let U denote union. Let I denote intersection.
C=(S1 U S2) represents the union of S1 and S2, called the cat set.
D=(S3 U S4) represents the union of S3 and S4, called dog set.
However (C I D) is not an empty set. In other words, there are neurons that fire for both cat and dog. A neurobiologist stumbles on one neuron at a time, in C or D. Not every one will do.

Do I explain clearly? Maybe you want to refute some of my above reasons.

-John



On 6/30/11 1:03 PM, Walter J Freeman wrote:

Experimentally, this is improbable.

> In other words, there are neurons that fire for both cat and dog.

Yes, but, not as members of sets.

-walter


On 6/30/11 2:36 PM, Juyang Weng wrote:

> Experimentally, this is improbable.

This is the normal behavior of cells. This is because the skull is closed during development.
No human teacher's hand is allowed inside the skull to supervise any cell to fire when, and only when,
a concept is present. Note the "only when" part.

I have attached a paper about the characteristics of emergent representation in the biological brain, considering the skull-closure fact.
It will appear in IEEE Transactions on Autonomous Mental Development. Let me know your comments. The following is a summary by an anonymous reviewer of that paper:

"
Without attempting to be in any way exhaustive, I would note the following key messages conveyed by the paper, focussing in particular on the clarifications that have been added to the current version:

- The external nature of symbolic representations.

- The need to deal incrementally with new external environments through real-time autonomously generated internal modal representations.

- The importance of modelling brain development.

- The now-clear distinction between emergent and symbolic representations.

- The differentiation between connectionist and emergent representations.

- The argument symbolic representations are open to ‘hand-crafting’ while emergent representations are not.

- The identification of the brain with the central nervous system.

- The external nature of the sensory and motor aspects of the brain.

- The consensual nature of the interpretation of sensations in human communication.

- The necessity for emergent representations to be based on self-generated states without imposing meaning or specifying boundaries on the meaning on these states.

- The transfer of temporal associations by the brain.

- The accumulation of spatiotemporal attention skills based on experience of the real physical world.

- The back-projection of signals, not errors, in descending connections.

- The importance of the developmental program and top-down attention for understanding human and artificial intelligence.
"

-John



On 7/1/11 3:08 AM, Walter J Freeman wrote:
Dear John,

It is not surprising that we have disjunct understandings of how brains work, you from a background in IT and electronics, I from neurology, histology, and experimental electrophysiology. I surmise that your impressive comprehension of neuroscience comes from the papers you cite in the article you sent me. I view this knowledge as comparable to an early 19th Century map of Africa, outlining the coasts (input and output) with the continent unknown.

The interior is where I have been experimenting for half a century. Some 30 years ago I concluded that brains don't make representations, either emergent or symbolic; they create meanings by constructing hypotheses that they test by the action-perception cycle. Hence the 'closed skull' you refer to (I prefer the 'blood-brain barrier' in physiology, and the 'solipsistic barrier' in philosophy). The list of topics from the reviewer of your paper indicates to me that you and I are addressing the same aspects of cognition and behavior, and that we hold similar concerns, but our models with which we implement our understandings take us in very different directions. My data base carries me a long distance from yours.

I respect your position but regretfully conclude that we are too far distant to communicate effectively within the limitations of our backgrounds and the time available for private discourse. Perhaps I'm wrong on that. I hope so.

Respectfully,

Walter


On 7/1/11 10:13 AM, Juyang Weng wrote:
The main source of my model is not just an early 19th Century map of Africa. It has considered:

- Biology: genomic equivalence, cell migration, cell differentiation, cell growth, and cell signaling.
From biology, we know that each cell is AUTONOMOUS. Any computational model (e.g., "concept cells") must explain how the concept can form through a developmental processing where every cell is AUTONOMOUS. E.g., how does a cell "know" concept?
- Neuroscience: of biological papers are mainly in neuro-anatomy, papers about areal responses (V1, V2, V3, IT, MT, PP, Frontal Cortex, motor areas, etc.), development-and-patterning, and plasticity.

> Some 30 years ago I concluded that brains don't make representations, either emergent or symbolic; they create meanings by constructing hypotheses that they test by the action-perception cycle.

That is a reasonable hypothesis, as the initial synapses are from spontaneous signals. The central issue is how they test by the action-perception cycle. My emergent
model is about how they test by the action-perception cycle AND adapt.

> you and I are addressing the same aspects of cognition and behavior, and that we hold similar concerns,

Great. That is why we are chatting.

> but our models with which we implement our understandings take us in very different directions.

This is natural and that is why we should communicate.

> My data base carries me a long distance from yours.

Please let me know a few examples, as I am interested.

> we are too far distant to communicate effectively within the limitations of our backgrounds

I agree. From the information source that I used to construct and experimentally verify our 5-chunk brain-mind model, I think that one needs to have sufficient
background in at least 6 disciplines: biology, neuroscience, psychology (including cognitive science), computer science, electrical engineering and mathematics.

The attached short text discusses "why 6 disciplines?", which echos your views. What do you think?

-John


On 7/4/11 12:16 PM, Walter J Freeman wrote:
Dear John,

I agree on the need for transdisciplinary action to test multidisciplinary hypotheses and adopt or adapt to failures of prediction, though I will insist on inclusion with your 6 of 1 more: philosophy. However, one cannot be expert in all subfields of these disciplines, so one must select with great care and cultivate only those areas that are most relevant.

Examples:

Biology: Yes, evolution is crucial, but not the evolution of logic. The foundations for intention, attention, memory, recall, recognition, navigation, and the action-perception-adaptation-orientation cycle were laid 400 million years ago in olfaction, not vision. The other distance senses (hearing, seeing, feeling) were added eons later with appropriate adaptations, while retaining and elaborating the same code. The commonality of code is essential for Gestalt formation, which follows perception and precedes prediction. Allocortex precedes neocortex.

Neuroscience: Cortex is bistable, having 2 phases, one gas-like, one liquid-like. Engineers have been seduced by the Golgi stain into their focus on neural networks. Indeed networks have their place for sensory pre-processing and motor particularization, but they don't scale up to billions of neurons and trillions of synapses. The receiving phase is particulate, infinite dimensional, Gaussian noise that is accessed with action potentials; the transmitting phase is a continuum with an order parameter that is accessed with dendritic potentials.

Psychology: The place to begin to understand higher cognition is the process of conversion of sensation to perception. This requires a phase transition comparable to condensation of a gas to a liquid followed by evaporation. The sensory phase requires pulses, bits, point processes; the perceptual phase requires pulse densities and wave densities. I dare say that the great majority of our colleagues are unaware of the existence of this information superhighway, by which every cortical neuron participates directly in every cognitive operation. Once this insight is achieved, decomposition is the next step.

Computer science: This is a tool, not a model. Let not the tail wag the dog. John von Neumann declared in his posthumous "The Computer and the Brain" that the language of the brain is not mathematics. Claude Shannon stated emphatically that his information theory could not serve to describe brain function. Paradoxically the mathematics of random noise devised by Shannon's contemporary at Bell Labs, Stephen Rice, is highly relevant to brain dynamics, though it is largely unrecognized.

Electrical Engineering: I have had close contacts with EECS departments for 40 years, particularly with Eugene Wong, Otto Smith, Lotfi Zadeh, Leon Chua, and Jose Principe (who nominated me for IEEE Fellow). From my knowledge of cortical electrophysiology and the use of multivariate statistics and nonlinear regression for curve fitting I can assure you that expert knowledge of piece-wise linearization and root locus techniques in feedback control theory is far more relevant to the dynamics of the transmitting phase than sparse coding and symbolic AI.

Mathematics: Yes, we do need 'a new kind of mathematics', and I have contributed to that goal. (My Erdos number is 2). The Hungarian Academician János Tóth wrote me in regard to the neural mechanism by which visual, auditory, somatic and olfactory signals are integrated into Gestalts:
At 9:35 AM +0200 6/26/11, János TÓTH wrote:
If you meant (i+k) x (j+l), then you again concatenate, this time two
matrices, in a very special way: the given matrices will be situated
as two squares of the same color on a chess table. But there is no
overlap here. There will be an overlap, however, if you force them to
take place in a smallar matrix then (i+k) x (j+l). This is an
absolutely new operation, nonstandard, but if it is important to
describe your experimental arrangement, you have the right to
introduce it.

It is important. The chess board is a useful model for an idealized cerebral cortex, as the starting point, but it must quickly be made to model neural integration by forcing overlap. We have done this effortlessly in Matlab code, but the real mathematics should be done by an expert, who I hope will become interested enough to collaborate, as I have persuaded experts in physics, mathematics and engineering to explore avenues I've opened.

Philosophy: Metaphysics is in disrepute among biologists and engineers. As the result there is a lack of sophistication leading to unacknowledged assumptions and widespread use of terms like 'information processing' and 'neural code' that are not recognized for what they are: metaphors. The terms are thin ice that enable users to skate over the depths of our ignorance. Most importantly, brains do not take in forms. As you wrote, the teacher cannot penetrate the skull. The two major philosophers who have promulgated this insight are Aquinas and the Buddha. For me it is a fascinating adventure to cull through the historical records and trace the connection through Arabic scholars of medieval Baghdad.

Parenthetically, my colleague Robert Kozma and I were the first to use the term 'intentional robotics', owing to my background in philosophy.

You ask, what do I think? I think that you model what brains do, and I model how brains do what they do. I think that your work well exceeds mine in your range and complexity of modeling human cognition, and that my work at present deals only with the pre-symbolic operations of the action-perception-adaptation-orientation cycle that we share with other animals, but it reveals processes undreamt of by engineers that only an experienced neurobiologist could unearth, and only a quantum field theorist can model (I don't mean quantum mechanics as practiced by the likes of Hammeroff and Penrose).

I think that my model might well serve as an interface between your symbolic dynamics and biological brain dynamics, if and when you relinquish your assumptions that brains make representations and use numbers at the level of sensation-perception. To claim otherwise is like saying the eye does a Fourier transform. It does not. It refracts. The physiologist and engineer compute the FT. I would ask you to concede von Neumann's insight:
"Thus the outward forms of our mathematics are not absolutely relevant from the point of view of evaluating what the mathematical or logical language truly used by the central nervous system is. ...
It is characterized by less logical and arithmetical depth than what we are normally used to. ...
Whatever the system is, it cannot fail to differ considerably from what we consciously and explicitly consider as mathematics."
John von Neumann (1958) The Computer and the Brain pp. 81-82

Now what do you think?

Walter

Principe, J.C., Tavares, V.G., Harris, J.G. and Freeman, W.J. (2001) Design and implementation of a biologically realistic olfactory cortex in analog VLSI. Proceedings IEEE 89: 1030-1051.

Ohl, F.W., Scheich, H., Freeman, W.J. (2001) Change in pattern of ongoing cortical activity with auditory category learning. Nature 412: 733-736.

Freeman WJ, Breakspear M [2007] Scale-free neocortical dynamics. Scholarpedia, 2(2): 1357.
http://www.scholarpedia.org/article/Scale-free_neocortical_dynamics

Freeman WJ [2007] Hilbert transform for brain waves. Scholarpedia 2(1): 1338.
http://www.scholarpedia.org/article/Hilbert_transform_for brain_waves

Freeman WJ [2007] Intentionality. Scholarpedia 2(2):1337
http://www.scholarpedia.org/article/Intentionality
Freeman, WJ, Vitiello G (2007) The dissipative quantum model of brain and laboratory observations. Electronic J Theoretical Physics 4, 1-18.
http://dx.doi.org/10.1016/j.plrev.2006.02.001

Kozma R, Huntsberger T, Aghazarian H, Tunstel E, Ilin R, Freeman WJ [2008] Implementing intentional robotics principles using SSR2K platform. Advanced Robotics 22(12): 1309-1327.

*Freeman WJ, Erwin H [2008] Freeman K-set. Scholarpedia, 3(2): 3238
http://www.scholarpedia.org/article/Freeman_K-set

Freeman WJ [2008] A pseudo-equilibrium thermodynamic model of information processing in nonlinear brain dynamics. Neural Networks 21: 257-265. http://repositories.cdlib.org/postprints/2781

Kozma R, Freeman WJ [2009] The KIV model of intentional dynamics and decision making. Neural Networks 22(3): 277-285. doi:10.1016/j.neunet.2009.03.019

Freeman WJ [2009] Deep analysis of perception through dynamic structures that emerge in cortical activity from self-regulated noise. Cognitive Neurodynamics 3(1): 105-116.
http://www.springerlink.com/content/v375t5l4t065m48q/
http://repositories.cdlib.org/postprints/3387
Freeman WJ, Kozma R, Bollobás B, Riordan O [2009] Chapter 7. *Scale-free cortical planar network, in: Handbook of Large-Scale Random Networks. Series: Bolyai Mathematical Studies, Bollobás B, Kozma R, Miklös D (Eds.), New York: Springer, Vol. 18, pp. 277-324.
http://www.springer.com/math/numbers/book/978-3-540-69394-9

*Freeman WJ [2009] The neurobiological infrastructure of natural computing: Intentionality. J New Math Natural Computing (NMNC) 5(1): 19-29. Special Issue on: Neurodynamics and Cognition in Recognition of W J Freeman (March 2009 issue) Kozma R, Caulfield HJ (eds.) http://repositories.cdlib.org/postprints/3378

Ramon C, Freeman WJ, Holmes MD, Ishimaru A, Haueisen J, Schimpf PH, Resvanian E (2009) Similarities between simulated spatial spectra of scalp EEG, MEG and structural MRI. Brain Topography 22:191-196.

*Freeman WJ, Vitiello G [2010] Vortices in brain waves. International Journal Modern Physics B, Vol: 24 (17): pp. 3269-3295. http://dx.doi.org/10.1142/S0217979210056025

--
Walter J Freeman MD, Professor of the Graduate School
Life Fellow, IEEE
Department of Molecular & Cell Biology
Division of Neurobiology, Donner 101
University of California at Berkeley
Berkeley CA 94720-3206 USA
Tel 510-642-4220 Fax 510-642-4146
dfreeman@berkeley.edu
http://soma.berkeley.edu



On 7/8/11 4:20 AM, Asim Roy wrote:
Hi All,

This is a restricted list. Please let me know if you want to be dropped from this list or want to add someone else to this discussion.

Attached is a writeup on concept cells, their nature and what they mean. It’s all based on the work by Christof Koch and others with epileptic patients at UCLA. Based on their recent experiments relating concept neurons to consciousness, it appears that a single spiking neuron can actually have meaning and is at the cognitive level. The writeup is short. Take a look and let me know if you would vehemently argue against such an inference. The conventional wisdom is that a single spiking neuron is too low level and at the noncognitive level, that its signal in isolation doesn’t “mean” anything, even though they might be abstracting information in some form.

And let me know if I have overreached in any of the other characterizations of concept cells. I think Christof and his group have some major discoveries.

Best,
Asim Roy
Arizona State University
www.lifeboat.com/ex/bios.asim.roy


On 7/8/11 8:57 AM, Juyang Weng wrote:
> Computer science: This is a tool, not a model.

Walter, the way the brain thinks is almost impossible to understand without the automata theory in computer science. Many people like Steven Harnad stated
that the brain network must do arbitrary re-combinations like a computer.
This is a major reason why neural network researchers still could not figure out a network can reason and think, at least re-combination on the fly!
Why my model can do it because I have a background in computer science. Automata theory is a model, but must do it in a brain-like network. Interesting?

-John


On 7/8/11 10:18 AM, Asim Roy wrote:
To those in the IJCNN workshop,

Would love to get some feedback on the idea that a single spiking neuron can be at the cognitive level, its signal can have meaning. Is that stretching the results of these experiments too far? Or does it sound like a reasonable inference?

Asim



On 7/8/11 11:02 AM, Harry Erwin wrote:
You're asking whether a single spike can be amplified, either into multiple parallel spikes or into a burst of spikes. The answer is 'certainly'. A single spike is usually amplified into the release of neurotransmitter at multiple synapses. Some cells in the auditory brain stem specialise in doing that very reliably. Look at inner hair cells, bushy and globular cells and the end bulb of Held. Also look at the role of calcium 'spikes' in making cells sensitive to afference during narrow time windows so that they respond to stimuli with a burst of spikes.




On 7/8/11 11:07 AM, Juyang Weng wrote:
Dear Asim,

I applaud that you try to move ahead with your academic beliefs. This kind of academic discussion is beneficial for improving our understanding of the brain-mind --- probably one of most complex object in nature.

However, as I raised, the large community related to the brain-mind subject is large, and it includes at least 6 disciplines (biology, neuroscience, psychology, computer science, electrical engineering, and mathematics), if you allow me to leave out philosophy, social science, linguistics, politics, laws, etc.

The large community is too fragmented, where each researcher's brain does not have sufficient knowledge about other disciplines to understand the brain-mind.

However, the human race collectively has already enough information to put the grand picture of the brain-mind jigsaw puzzle together! This conclusion is based on my (5+1) chunk model that has used a vast amount of evidence that humans have discovered (in these 6 disciplines and more). The 5-chunk model has been presented in a session in IJCNN 2010, which you attended. The addition chunk is brain modulation, which will appear in IJCNN 2011. Hence 5+1 chunks.

The following is in the conclusion of my manuscript "A (5+1)-Chunk Theory of the Brain-Mind Networks" currently under review, which is the archival version of Weng IJCNN 2010. The paragraph below is in its conclusion:

"Therefore, due to the highly complex nature of motor information (in $\v_z$) as emergent temporal context (e.g., an event), the DN model predicts that there is no symbolic extra-body meaning (e.g., edge or face) that exactly corresponds to the role of any neuron or any collection of neurons in the biological brain, regardless how the symbolic extra-body meaning is expressed in a human language. The role of a biological neuron does not exactly correspond to, but can only roughly correlate with, an extra-body meaning."

Your "concept neuron" does not have a basis, if it cannot explain the first chunk: development, let alone other chunks.

>From Weng IJCNN 2010: "The `development' chunk has task-nonspecificity, emergent representation, and skull-closedness as necessary conditions for the brain."

When the skull is closed, how a concept in the mind of a teacher gets into exactly one neuron inside the child's brain, so that the neuron fires if and only if the concept is present in the extra-body environment? Note the ``only if'' part, which is what Jay and others are reminding you for. Jay used "about 150 other things" a neuron may respond to. The brain skull is closed for your goal of "autonomous learning". The developmental program (DP) inside the closed brain MUST fully autonomously learn. The DP has no "idea" about what an extra-body concept is.

Let us define a concept neuron precisely:

"A concept neuron of concept C is such a neuron inside the brain that fires if and only if the concept C about the extra-body environment is true, where the extra-body environment contains an open number of possible natural objects". Note, not an intra-body concept (such as "neuron", or "the left arm"), as the brain has direct access to the body.

In the above sense, there is no conceptual neuron anywhere in the brain. I can bet any amount of money on it with anybody. To be more doable, let us say that
there is no concept neurons in IT. We can formally set a bet, using a half of the amount of money to award to a team that can provide a satisfactory proof or disproof.

Best,

-John


On 7/8/11 11:13 AM, Juyang Weng wrote:
This answer has shown how far Asim's question can be misinterpreted and misunderstood.

-John



On 7/8/11 11:02 AM, Harry Erwin wrote:

You're asking whether a single spike can be amplified, either into multiple parallel spikes or into a burst of spikes. The answer is 'certainly'. A single spike is usually amplified into the release of neurotransmitter at multiple synapses. Some cells in the auditory brain stem specialise in doing that very reliably. Look at inner hair cells, bushy and globular cells and the end bulb of Held. Also look at the role of calcium 'spikes' in making cells sensitive to afference during narrow time windows so that they respond to stimuli with a burst of spikes.


On 7/8/11 11:13 AM, Yoonsuck Choe wrote:

Hi Asim and colleagues,

I can't say yes or no for all neurons, but at least there are certain
classes of neurons that could carry cognitive (or complex) meaning.

The class that I have in mind is none other than that of mirror neurons. Mirror neurons respond to visually perceived gestures, and they also activate when the same gesture is performed by the same animal. In this sense, the representation (of gesture) have an exact meaning (congruent performance).

We have previously investigated meaning of single neural spikes, and it led us to the conclusion that action is necessary and passive perception cannot lead to inferred meaning. We have also found that action patterns that maintain internal state invariance can provide meaning for single
spikes.

Here's a really brief 2-page summary of this idea, to be presented at
an AAAI workshop:

http://faculty.cs.tamu.edu/choe/ftp/publications/choe.aaai11.pdf

A more extensive coverage can be found in

http://faculty.cs.tamu.edu/choe/ftp/publications/choe.aaai06.pdf
http://faculty.cs.tamu.edu/choe/ftp/publications/choe.ijhr07-reprint.pdf

Yoonsuck
choe@tamu.edu


On 7/8/11 11:36 AM, Asim Roy wrote:
Hi John,

There is no claim of “if and only if” type of concept neurons anywhere. Sure 150 other neurons may fire. Take a look at this paper, it is really interesting. Really nice experiment.

Cerf, M., Thiruvengadam, N., Mormann, F., Kraskov, A., Quian-Quiroga, R., Koch, C. & Fried, I. (2010). Online, voluntary control of human temporal lobe neurons. Nature, 467, 7319, 1104-1108.

Asim



On 7/8/11 11:42 AM, Walter J Freeman wrote:
> Computer science: This is a tool, not a model.

Walter, the way the brain thinks is almost impossible to understand without the automata theory in computer science. Many people like Steven Harnad stated
that the brain network must do arbitrary re-combinations like a computer.
This is a major reason why neural network researchers still could not figure out a network can reason and think, at least re-combination on the fly!
Why my model can do it because I have a background in computer science. Automata theory is a model, but must do it in a brain-like network. Interesting?
-John

Very interesting. The race is on. I hope I live to see which way the wind blows.

Walter



On 7/8/11 2:03 PM, Juyang Weng wrote:
Asim,

If there is no "if and only if" condition for a cell, then there is no concept cell. It is not "150 other neurons may fire" when Jennifer Aniston is present. It is the same cell that responded to Jennifer Aniston will respond to about 150 other things, including "pumpkin" as Ali said.

I read that Christof's Nature paper as soon as it appeared. Christof told me about it in Singapore when we were at Decade of Minds conference last year, but it was in the media embargo status before it appeared. In my view, the "Jennifer Aniston" phenomenon in IT is basically the same as "an orientation cell" in V1, they are all specific for the same computational principles.

Sorry, Christof, it was indeed an interesting article. But Nature editors could not appreciate our theory that explains what you discovered and we demonstrated in our brain-simulation computer experiments. They finally said that our Nature submission was not "competitive enough", apparently beaten by your paper. My sincere congratulations!

The mind (via motor areas in our model) controls an IT cell and a V1 cell, in the same way. In our Where-What Network (WWN) 3, the mind controls all learned concepts, not just a type cell (Jennifer Aniston, pumpkin, or 67-degree-edge), but also location (what is at row-2-and-column-3?), scale (find an object of interest of a 30x30-pixel scale, but do not care of an 13x13-pixel scale). In a WWN, each cell will respond to about hundred of other things, but each motor cell learns a type concept (Jennifer Aniston, pumpkin, or 67-degree-edge in the type motor area, e.g., vocal tract) or a location concept (row-2-and-column-3 in the location motor area, e.g., muscle neurons for arm pointing). However, we use a single neuron to represent a type concept or a location concept simply because our teacher-taught language is overly simple. A muscle neuron in the vocal tract will respond to may names --- Jennifer Aniston, pumpkin, or 67-degree-edge. A muscle in my left arm will fire when my arm points to row-2-and-column-3, when I dance, when I write ...

There is no concept cell in the brain, period. What is interesting is that a WWN network can learn all kinds of concepts, and it performs perfectly with zero error, as long as it has enough neuronal resource and good training experience. The brain-mind in the 5-chunk scale has already appeared, but Science and Nature editors could not understand it (of course! Do they have knowledge in 6 disciplines?). A reviewer commented "the authors are trying to do in a mid-length journal article what Edelman tried to do in 3 books."

Asim, you told me a few times in a year that you will read Weng IJCNN 2010. That is why you are still taking about "concept cells". I am sorry, as the material in that paper is extremely difficult to understand, as it tells many things that run counter to human intuitions --- one of the intuitions is "concept cell".

-John


On 7/8/11 2:15 PM, Juyang Weng wrote:
Yoonsuck,

all motor neurons in WWN are mirror neurons. I did not mention "mirror neuron" in that paper because of a lack of space. Seeing a hand in WWN
will cause WWN to pronounce "hand" in TM and its "hand" location neurons fire too. However, a mirror neuron is not a concept neuron due to the detailed reasons in my last email. In my humble opinion, to understand mind-brain computationally, we need rigorous definitions, not just vague terms.

-John


On 7/8/11 2:49 PM, Asim Roy wrote:
Hi John,

In that note I sent, I tried to characterize concept cells (Christof’s term) based on the findings of several studies by Christof and his associates. I did not try to define it, just tried to summarize the properties of these cells they found and described in their papers. Perhaps you have a certain definition of concept cells in your mind that is in conflict with what they found in their studies. It might be appropriate to use different terms here.

I am glad you are using the term “control” with regard to the mind/brain and how it works.

Asim



On 7/8/11 2:56 PM, Asim Roy wrote:
Hi Yoonsuck,

I will take a look at these. From what you describe, I think mirror
neurons are at the cognitive level. Are there experimental studies on
mirror neurons that I can reference?

Asim




On 7/8/11 8:22 PM, Yoonsuck Choe wrote:

Hi Asim,

Thanks! The 2-page summary would be enough to get you started.

Mirror neurons were first discovered experimentally, and their theoretical implications have build up quite a bit over time. Relevant papers would be (these are the earliest ones):

V. Gallese et al., Action recognition in the premotor cortex. Brain 119 (1996), pp. 593-609.

G. Rizzolatti et al., Premotor cortex and the recognition of motor actions. Cognit. Brain Res. 3 (1996), pp. 131-141.

G. Rizzolatti and M.A. Arbib, Language within our grasp. Trends Neurosci. 21 (1998), pp. 188-194

Also, take a look at VS Ramachandran's essay on Edge.org:

Mirror neurons and imitation learning as the driving force behind “the great leap forward” in human evolution

http://www.edge.org/3rd_culture/ramachandran/ramachandran_index.html

Yoonsuck
choe@tamu.edu


On 7/8/11 8:50 PM, Yoonsuck Choe wrote:

Dear John,

Thanks for your comments. Yes, I recall that your What-Where-Network neurons have the mirror property. I'll have to revisit your email to see the finer points about concept cells that I didn't grasp in my first reading. I'll let you know if I have further questions.

BTW, can you clarify a bit (or point to your published work) about "rigorous definitions"? I am familiar with your 6-tuple representation,
forward/backward/lateral projections, etc. I guess you are talking about
definitions like these?

To make progress in understanding the brain, I think we need to first reevaluate the research questions. We need to pose the questions from the perspective of the brain itself: Are we making the same assumptions based on which the brain operates? For example, neurons in the brain operate solely based on spikes: They are never given the actual environmental input. So, methods like reverse correlation cannot be used by the neurons
themselves, although such methods are powerful tools for neuroscientists.

Yoonsuck
choe@tamu.edu

Yoonsuck Choe, Ph.D.


On 7/9/11 12:49 AM, Juyang Weng wrote:
Dear Yoonsuck,

My definition was in an earlier email dated 7/8/11 11:07 AM

"A concept neuron of concept C is such a neuron inside the brain that fires if and only if the concept C about the extra-body environment is true, where the extra-body environment contains an open number of possible natural objects". Note, not an intra-body concept (such as "neuron", or "the left arm"), as the brain has direct access to the body.

Note the "only if" part. Without it, the definition is useless. Asim does not like the "only if" part. If a neuron responds to "Jennifer Aniston", "pumpkins" and about 150 other things, do we want to call it concept cell?

-John


On 7/9/11 12:51 AM, Juyang Weng wrote:
Yoonsuck,

any mirror neuron is not a concept cell, as it does not satisfy the
"only if" part in my definition of a concept cell.

-John



On 7/9/11 1:07 AM, Juyang Weng wrote:
Asim,

Vague terms are not going to give us any useful discussion.

When a subject is immature, researchers use vague terms to describe individual phenomena.

However, vague terms are not sufficient for computational modeling. A computational process demands very precise terms so that the process can be described in mathematics, implemented, and verified. Suppose that you do not use the "only if" part. If a neuron responds to "Jennifer Aniston", "pumpkins" and about 150 other things, do you want to call it "Jennifer Aniston" concept cell?

The term "control" I used for motor neuron to control top-down attention is only an intuitive term. It is very different from any engineering sense of controller, as it is not the only cell that affects top-down attention. In other words, it is almost never in total control.

-John



On 7/9/11 2:12 AM, Harry Erwin wrote:
The concept should be fairly common or the cell will remain quiet for long periods of time and be vulnerable to pruning.



On 8 Jul 2011, at 09:49, Juyang Weng wrote:
Dear Yoonsuck,

My definition was in an earlier email dated 7/8/11 11:07 AM

"A concept neuron of concept C is such a neuron inside the brain that fires if and only if the concept C about the extra-body environment is true, where the extra-body environment contains an open number of possible natural objects". Note, not an intra-body concept (such as "neuron", or "the left arm"), as the brain has direct access to the body.

Note the "only if" part. Without it, the definition is useless. Asim does not like the "only if" part. If a neuron responds to "Jennifer Aniston", "pumpkins" and about 150 other things, do we want to call it concept cell?

-John



On 7/9/11 4:53 AM, Yoonsuck Choe wrote:

Dear John,

Thanks. I saw that but didn't make the connection.

Yoonsuck
choe@tamu.edu

Yoonsuck Choe, Ph.D.



On 7/9/11 6:03 AM, Asim Roy wrote:
I invite John Collins to join this discussion. He is a professor of physics at Penn State (http://www.phys.psu.edu/~collins/)and has looked at some of the same data on concept cells. John, would love to get your insights.

Please let me know if you want to drop out of this discussion. This can go on for a while.

Asim



On 7/9/11 10:01 AM, Juyang Weng wrote:
One view: "A neuron actually has no lifespan. They die during an accident, disease or when the brain dies."
If neuron does not fire for long periods, will it die? In our model, it will be there for long time to serve as the needed long-term memory.
However, it will be recruited and updated if similar events occur, causing update-induced forgetting about the old details.

-John


On 7/10/11 10:09 AM, Asim Roy wrote:
Here’s my read on the experiments conducted by Christof and his team.

There is an explicit attempt by the brain to abstract high-level information in order to simplify and automate certain tasks. For example, they mention dedicated cells being created within days to recognize the researchers doing the experiments. Koch (2011) nicely generalized from their experiments: “Every time you encounter a particular person or object, a similar pattern of spiking neurons is generated in higher-order cortical regions. The networks in the medial temporal lobe recognize such repeating patterns and dedicate specific neurons to them. ….Conversely, you do not have concept cells for things you rarely encounter, such as the barista who just handed you a nonfat chai latte tea.” There is certainly an efficiency aspect to such abstractions. They can avoid the work involved in reinterpreting the same pattern over and over again. That task is automated and simplified with the dedicated neurons. And such high level abstraction should also mean not confusing Jennifer Aniston with a pumpkin. It appears that these dedicated cells are all about simplification, automation and computational efficiency. To me, they seem like good properties from an engineering point of view.

Jeff Bowers has an article, I believe in Connection Science, where he explains how a unit (a single neuron) can code for (represent??) say the word “Harry,” but may also respond to similar words like “Barry,” “Marry” and so on. I guess by his argument, a neuron can represent an abstract concept (“the basketball players,” “the landmarks,” “the Jedi of Star Wars”), but possibly respond in some other cases. They need be “only if” response type.

The more interesting question is whether a single spiking neuron can have meaning. In other words, can it be at the cognitive level. Of course, meaning requires rigorous definition and if anyone knows where we can find one, that would be much appreciated. If one looks at cognitive science models of semantics, they are generally constructed to learn semantics of the following type: when you see a canary, you know that it’s a bird, it can fly, that it’s an animal and so on. The models learn these types of relationships. The recent experiments by Cerf, Koch and others are of a somewhat different nature. In their experiments, the patients were looking at two superimposed images and they were asked to “think” about one of those images. By tracking the firings of two particular spiking neurons, each one matched to one of the images (there were two other neurons matched to two other images not present in the display), the researchers could tell if a patient was actually “thinking” of one image or the other. In other words, one spiking neuron can effectively track a particular thought. And such “thought” tracking was verified in hundreds of experiments with twelve different patients. Does that mean that spikes of some dedicated neurons have meaning? Would this qualify as an adequate test for “meaning?”

Look forward to your views on the question of a single spiking neuron being at the cognitive level and having meaning. These are definitely highly debatable issues. Wondering if there are other experiments like these. Yoonsuck did mention his work with single neurons and mirror neurons having meaning.

Asim


On 7/11/11 10:33 AM, Juyang Weng wrote:
> “Every time you encounter a particular person or object, a similar pattern of spiking neurons is generated in higher-order cortical regions. The networks in the medial temporal lobe recognize such repeating patterns and dedicate specific neurons to them. ….Conversely, you do not have concept cells for things you rarely encounter, such as the barista who just handed you a nonfat chai latte tea.”

Christof did not say that these "specific neurons" do not respond to many other things. There is a scope issue in "repeating patterns": repeated patterns of what scope in the brain? In Weng IJCNN 2010, I criticized symbolic representation for the brain. "Because of the symbol use, the modelers use “skull-open” approaches — the holistically aware central controller is the outside human designer." I stated that Treisman's master map idea is wrong, although it has been used by well known neuroscientists like Van Essen. I wrote "(1) There exists no master map in the brain. (2) A receptive field is in general dynamic, not appearance kept or even topology kept."

Basically, there is no central controller in the brain (remember your controller-like view?). After mitosis, each neuron does not "know" its role. It almost "blindly" connects with other neurons based on its cellular properties. The role of each neuron is jointly determined by its properties and the neurons that it connects with.
Therefore, the role cannot be based on extra-body concepts, only based on intra-body concepts.

A deeper reason is that the skull is closed during development. Abstract meanings like a "concept" about the extra-body concept cannot get directly to each individual neuron. Interestingly, we have a 5-chunk brain-mind model that does not need such an "skull-open" approach, but it can do recombination of
extra-body concepts on the fly like a digital computer. To understand how, one needs to learn computer science, especially automata theory. Many neural network researchers do not have a background in computer science.

> The more interesting question is whether a single spiking neuron can have meaning.

Of course, each single spiking neuron has specific meanings, but in terms of intra-body concepts only (e.g., a neuron in the IT area which connects with some neurons in V1, V2, V3, V4, LIP, the ventral frontal area, and the pre-motor area). However, these meanings do not exactly correspond to any extra-body concepts (not Jennifer Aniston specific, not pumpkin specific).

In an TAMD accepted paper "Symbolic Models and Emergent Models, A Review", I wrote: "I argue through this review that the brain does not seem to contain any computer symbol at all in its internal representations for the extra-body environment (i.e., outside the body)."

Any model about the brain must be explainable by known biological mechanisms. Without biological supports, any model is subject to major flaws. Treisman's master map is an example.

-John



On 7/11/11 3:20 PM, Asim Roy wrote:
Hi John,

The controller issue is a side issue here. But since you bring it up, there is no claim in my paper that there is a central controller. The claim is that parts of the brain control other parts. That’s all. In that paper, we analyzed standard connectionist learning algorithms and showed that they depend on outside control and intervention in order to learn. And I believe your algorithms are no exception. We have discussed this. But let’s not argue on the controller issue here. The main issue is whether a single spiking neuron can have meaning. Let’s focus on that.
There is no claim anywhere by anyone - not by Christof, not by me and not by anyone else - that a concept neuron will respond to only one thing. Jeffrey Bowers has been arguing about grandmother cells for a while and even he doesn’t claim that grandmother cells, if they exists, will respond to only one thing. He recently wrote a paper explaining that:
Jeffrey Bowers, “What is a grandmother cell? And how would you know if you found one?” Connection Science, Vol. 23, No. 2, June 2011, 91–95.

Here’s a quote from his paper in reference to the interactive activation (IA) model of McClelland and Rumelhart 1981 for word identification, which is a localist model with each output unit representing a particular word:

“Although each word unit (e.g. TRAP) selectively represents one word (e.g. trap), each word activates multiple word units and each word unit is activated by multiple words. For example, in response to the word trap, the TRAP unit will become most active, but the units for WRAP, TRIP, TRAM, TRAY, etc., will become active as well (by virtue of sharing three letters with trap). Nevertheless, the WRAP, TRIP, TRAM, and TRAY units play no role in representing the word trap. Indeed, removing all the units would have no impact on the representation of the word trap (the word trap would still be recognized when the activation of the TRAP unit passes some threshold, just as before).”

So “if and only if” is not an appropriate condition to impose on the concepts of grandmother cells or concept cells, given the input-output structure of these networks and the similarity of the patterns they have to recognize.

On single spiking neurons having meaning, you say:

“Of course, each single spiking neuron has specific meanings, but in terms of intra-body concepts only (e.g., a neuron in the IT area which connects with some neurons in V1, V2, V3, V4, LIP, the ventral frontal area, and the pre-motor area). However, these meanings do not exactly correspond to any extra-body concepts (not Jennifer Aniston specific, not pumpkin specific).”

So are you saying that a single spiking neuron can have “meaning” and encode “a concept,” but only of a certain type (intra-body)? And that these have the “if and only if” properties that you want? If so, that sounds great. Would greatly appreciate some references (from neuroscience??) that have definitely verified the “if and only if” nature of such spiking neurons. I think that would be very useful to our discussion.

Asim



On 7/12/11 11:17 PM, Juyang Weng wrote:
> The claim is that parts of the brain control other parts. That’s all.

Asim, our DN brain-mind model shows that no part of the brain is in total control of any other parts, only affects other areas' operations to some degree. Is it consistent to your term "control"?

> we analyzed standard connectionist learning algorithms and showed that they depend on outside control and intervention in order to learn. And I believe your algorithms are no exception.

Asim, absolutely not in my algorithms. You and I communicated about AMD before, but you were pointing to the resource parameters of our development program. Resource parameter is not "control", only regulation --- a well accepted term in biology. I think that any genome program must have some way to regulate the brain resource (e.g., the number of neurons in the brain). Otherwise, we will not say that the brain of a human adult has about 100 billions neurons.
In summary, regulate, not "control".

Note, I repeated used the term "skull closed", "task nonspecificity" for AMD. A human teacher is not in control of the Developmental Network (DN). The Developmental Program (DP) of the Developmental Networks is "fully autonomous" inside the closed skull.

> There is no claim anywhere by anyone - not by Christof, not by me and not by anyone else - that a concept neuron will respond to only one thing.

Then, concept of what?

> “Although each word unit (e.g. TRAP) selectively represents one word (e.g. trap),

According to the text that follows in Bowers 2011, the above term is inappropriate. No neuron is in a position to represent any one word, if the skull is closed. Bowers should read any of our DN papers to see that there is no need for any neuron to represent any word in order for a DN network to work well (for the training set: error-free, no iteration, immediate learning; for the test set: optimal in maximum likelihood).

> So “if and only if” is not an appropriate condition to impose on the concepts of grandmother cells or concept cells,

Then, can we simply drop the terms like "grandmother cells or concept cells"? They are misleading at least.

> So are you saying that a single spiking neuron can have “meaning” and encode “a concept,” but only of a certain type (intra-body)?

Intra-body concepts are well known in biological development, since each cell has direct communication with other cells inside the body and the brain (not directly with extra-body environments!) Within the body and the brain, a cell can migrate to a zone near a tissue and communicate with other cells in the tissue and then turn itself into a cell the same type (e.g., through "cell adhesion" or "cell contact"). In biology, it is called the fate (role, type, shape, connections, and other properties) of a cell. The fate of a cell is not determined when the cell is generated from a non-symmetric mitosis (called non-symmetric cell differentiation). The fate of a cell depends on (a) its differentiation during mitosis (early cell type), (b) what happens along its path of cell migration, and (c) the cellular environment after it has reached the destination (e.g., growth cone of each of its axons enables the axon to extend). Note: each neuron is a cell.

A short reference:
M. Sur and J. L. R. Rubenstein, "Patterning and Plasticity of the Cerebral Cortex, Science, 310 (805-810), 2005.

A more tutorial like reference (pp. 1019-1130, Fig. 54-7 on page 1071 is very intuitive):
E. R. Kandel and J. H. Schwartz and T. M. Jessell, Principles of Neural Science, 4th edition, McGraw-Hill, New York, 2000.

Asim: Do not get turned off by biology like many AI/CS/IT/Psy researchers! Although the biological material is very detailed, it is critical for us to understand what information each cell inside the the brain uses to migrate, connect and find its role! This is a kind of critical knowledge one needs if he truly wants to understand how the brain-mind works. Many cognitive scientists and psychologists (including many who publish in the Connection Science journal) do not bother to look into neuroscience, let alone biology! Isolation of individual disciplines is one of the ``Two Hindrances of Science" that I talked about in AMD Newsletter vol 8, no. 1, 2011: http://www.cse.msu.edu/amdtc/amdnl/AMDNL-V8-N1.pdf

-John


On 7/12/11 11:31 PM, Peter Foldiak wrote:

2. There is no claim anywhere by anyone - not by Christof, not by me and not by anyone else - that a concept neuron will respond to only one thing.

I would. Why not, if the "thing" is frequent enough.




On 7/12/11 11:42 PM, Christof Koch wrote:
I fully agree with Peter. It may well be possible. Showing this conclusively, i.e. that is responds to nothing else, is not practical though

;-)
----
Dr. Christof Koch



On 7/13/11 3:16 AM, Asim Roy wrote:
Hi John,

Define the terms “control,” “total control,” “regulate” and the various other ones you have used. Define them in a rigorous way like you did for “concept cells.” Your algorithms are based on a trial-and-error process and require outside intervention. I would rather leave the control issue out of this. We have been through this before. Let’s focus on concept cells.

Didn’t quite get answers to the questions on intra-body concept cells. Here they are again:

Do they have the “if and only if” property of a concept cell as per your definition? And is that “if and only if” property verified in biology? If so, what are the references for that verification.
If these intra-body concept cells have the “if and only if” property, do they then have “meaning?” If so, what kind of “meaning?” I think you have alluded to some before, but just want to clarify it once more.

Asim




On 7/13/11 3:41 AM, Asim Roy wrote:
I think the “simplicity, automation and computational efficiency” argument favors the creation of dedicated neurons for concepts that are encountered frequently. That’s essentially Peter’s frequency argument.

Back to my question: Do these concept cells have meaning? In Christof’s experiments, they could track down the firings of a particular neuron to a particular thought about a particular image, even though multiple images were on display in front of the patient. And the patient getting feedback through the enhancement of a particular image that he/she is thinking about, is that enough validation/verfication that a single spiking neuron can have meaning that one can relate to? In these experiments, the patient does get a confirmation of the thinking about a particular image.

Asim



On 7/13/11 4:00 AM, jeff bowers wrote:
Hi Christof, interesting that you would entertain the idea that a neuron (under some circumstances) might respond to only one “thing”. Does this mean that you think the neuroscience is at least compatible with the hypothesis of grandmother cells? What is your current take on claim that finding a “Jennifer Aniston” cell rules out the grandmother cell hypothesis? This is the basic conclusion in papers such as:
Quiroga, Q. R., Kreiman, G., Koch, C., & Fried, I. (2008). Sparse but not “grandmother-cell” coding in the medial temporal lobe. Trends in Cognitive Sciences, 12, 87–91.
Would be nice to have another voice saying that the ideal of grandmother cells is not totally absurd, and already ruled out by data (it is pretty lonely out there!).
Although a cell that responded to only one “thing” would seem to satisfy the grandmother cell hypothesis, I don’t think that is necessary. As noted in previous posts, grandmother cells might fire to related things to a diminished degree (like a simple cell tuned to vertical might respond to an edge a few degrees off). Orientation might be a bad example in some ways – but the idea that it is important to distinguish between what a neuron represents and what it responds to is summarized in the following short paper entitled: What is a grandmother cell? And how
would you know if you found one? (I noted it before, but I thought I would attach it).
Jeff
p.s. John, some people who publish in Connection Science pay attention to the neuroscience. Check out my Psychological Review paper that provides a detailed review of single cell recording data into this question, from single neuron in the visual system of the blowfly to hippocampus of humans. It might be all wrong, but plenty of neuroscience.
Bowers, J.S. (2009). On the biological plausibility of grandmother cells: Implications for neural network theories in psychology and neuroscience. Psychological Review, 116, 220-251. PDF



On 7/13/11 4:20 AM, Asim Roy wrote:
John Collins has an insightful note on concepts cells. I have requested him to post it on this list.

All, please let me know if you want to be dropped from this list or would like to add others.

By the way, there is nothing personal about these arguments. Scientists and academics are always argumentative until everything is settled and verified one way or the other. We argue hard, but we are all friends in the end.

Best,
Asim



On 7/13/11 10:50 AM, Juyang Weng wrote:
Hi Christof,

I fully understand the difficulties in brain response recording, as I was doing electrode and imaging recording from an animal brain in Sur's Lab. I am not saying that I know everything.

> Showing this conclusively, i.e. that is responds to nothing else, is not practical though

Yes, it seems practical. In fact, this has been done and reported many times in the literature. For example, Kobatake & Tanaka 1994 reported that a neuron that they selected from IT responded to (a) the face of a toy monkey, (b) two black dots above a horizontal black bar inside a gray disk (2-dots-above-bar-inside-circle), but not to (c) the same 2-dots-above-bar-inside-circle but in opposite contrast. The same neuron does not respond to 2-dots-inside-circle without bar, bar-inside-circle without 2 pots, and 2-dots-above-bar without circle. This study is consistent to our DN model: A neuron in IT does not detect an abstract extra-body concept (I guess it is "face" in the above case). Otherwise, it would have responded to (c).

Second, what each cell responds to depends on not only its bottom-up connections, but also its top-down connections. For example, Motter 1994 (J. of NS) reported that whether a V4 cell fires or not depends on not only whether its bottom-up receptive field has a red-diagonal-bar but also the current task goal. In this case, the task goal is related to the color of the fixation point outside the bottom-up receptive field of this neuron. Since whether the V4 cell fires depends on the current task goal and task goals are dynamic and open ended, we should not call this V4 cell as red-diagonal-bar concept cell. Otherwise, it would fire at the presentation of a red-diagonal-bar (which is not unique anyway per above paragraph) regardless the current task goal. This study is also consistent to our DN model: Each cell in the brain has a top-down connection part and that part can be related to open-ended temporal contexts and task goals. This makes no cell in IT (or any internal areas of the brain for that matter) to be specific to any extra-body concept.

Third, what each cell responds to also depends on what other cells (in the same area and in other areas) respond to. There are many brain plasticity studies. Elman 1997 "Rethinking Innateness" has a chapter that gives an excellent summary about such plasticity studies. For a "same area" case, removal of a section in a cortical area causes other remaining cells to take over the job that the removed cell took. For an "other area" case, Sur and coworkers 2003 (Science vol. 300, 2003) showed that MGN cells (hearing) takes over the visual processing task when cells in Superior colliculus were cut off.

Please allow me to I summarize my major points here:
(1) Each cell inside the brain has a bottom-up connection part, which does not respond to a unique sensory concept (because, e.g., the receptive field is dynamic!).
(2) Each cell inside the brain has a top-down connection part, which corresponds to rich and dynamic spatiotemporal contexts and task goals.
(3) Each cell inside the brain has a lateral connection part, which enables it to compete for roles with other connected cells.

Let us move ahead toward our exciting not-very-far future. I feel that such experimental neuroscience papers about cell behaviors are useful. However, such papers only report experimental phenomena, not causality. Papers about neuro-anatomy are also experimental, but give some reasons about cell behaviors. For deeper insight of neuro-anatomy, biological papers about brain development are useful to understand how such neuro-anatomy emerges, not statically specified by the genome.

-John



On 7/13/11 3:06 PM, Juyang Weng wrote:
1. For the parameter of resource (number of neurons in the brain Y), we do not have a trial-and-error process. We proved maximum likelihood for every number of parameter.

For AMD, humans intervene at the DP (Developmental Program, or the genome program) programming stage, before the conception, to short cut the extremely expensive and slow evolution.

In contrast, natural evolution is a process of trial-and-error. If a genome is not good enough (birth defects), the child cannot develop well to reach the mating stage. When he dies, his genome is lost from the pool of all genomes of the species. That is how a species evolves.

2.a Yes, each cell, in terms of intra-body concepts, has the "if and only if" property. For example, a cell in IT, in terms of (1) the set of neurons that it connects with at each time instant and (2) the weight vectors that it uses to connect them at each time instant, is quite unique inside the brain. Before, we used smooth representation like SOM (nearby neurons are similar in terms of intra-body concepts). We have realized that it is not necessarily best computationally and not what the biology does.

2.b Yes, each has a meaning in terms of the intra-body concepts if such concepts are detailed enough to make it unique (like an address system if the address is detailed enough). But, each neuron does not have an "if and only if" meaning in terms of extra-body concepts.

There seems no "Jennifer Aniston" specific cell if we consider the following points:
(a) Many cells fire responding to pasts or the whole image patch of "Jennifer Aniston".
(b) Each cell in (a) also responds to many other things, such as "pumpkins" and about 150 other things.
(c) At each time instant, there are many objects in the background, in addition to "Jennifer Aniston" (e.g., furniture or trees). Many neurons respond to such background objects. Whether a neuron fires depends on not only "Jennifer Aniston", but also other objects in the background, the spatiotemporal context in the past, and the task goals. The papers that I cited in the last emails support these points, and are consistent with our DN model.

-John



On 7/13/11 4:03 PM, Juyang Weng wrote:
Asim,

"Simplicity, automation, and computational efficiency" should not be the primary reasons for DP through evolution.

First of all, the most basic reason must be autonomous development --- from a single cell to develop into an adult cell through interactions with the environments. Three major requirements for autonomous development (also your goal of autonomous learning) are:
(a) Skull closure,
(b) Task nonspecificity (DP or the genome program cannot precisely predict what tasks and what environments the animal will deal with at each particular lifetime), and
(c) Emergent representations (all representation must emerge autonomously inside the closed skull, without requiring the parent to open the skull to implant extra-body concepts into specific neurons since the parent is too stupid for doing this impossible task)

-John



On 7/13/11 4:33 PM, Asim Roy wrote:
Hi John,

On the control issue, we need definitions for “control,” “total control” and “regulation.” Otherwise we will be beating around the bush. I have looked at your algorithms, although it’s been a while, and if I recall correctly, there are other trial-and-error processes.
On concept cells, you are accepting the fact that the brain is capable of creating “if and only if” concept cells, although only for intra-body concepts. And you are accepting the fact that those concept cells have “meaning,” in some undefined sense. Well, that’s a big step forward in the sense of acknowledging that such a capability exists in the brain. The findings by Christof and his team, that dedicated neurons are created to handle repeating patterns, are very much consistent with the capability that you acknowledge exists in the brain.
The Kobatake & Tanaka 1994 experiment that you cite to claim that extra-body concepts cannot be encoded in the same way as intra-body ones, is perhaps not tracking neurons over a longer period of time for the appropriate concept cells to be developed. Christof’s experiments allow time for repeating patterns to occur and then be encoded by dedicated neurons. Thus, occurrence of repeating patterns may be a pre-condition for dedicated neurons (concept cells) to be created by the brain. And I would guess that repeating patterns is the reason intra-body concepts are created also.
Asim



On 7/13/11 4:42 PM, Juyang Weng wrote:
> there are other trial-and-error processes

All the parameters we use in DN and WWN are reasonable in the genome. You need to let me know which is not, but you need to learn what genes can do in biology to be convinced. You need to teach yourself biology.

> And you are accepting the fact that those concept cells have “meaning,”

If and only if it is about intra-body concepts.

> The Kobatake & Tanaka 1994 experiment that you cite to claim that extra-body concepts cannot be encoded in the same way as intra-body ones, is perhaps not tracking neurons over a longer period of time for the appropriate concept cells to be developed.

Not true. Read the paper. They recorded long time.

-John



On 7/13/11 4:43 PM, Asim Roy wrote:
John,

In none of Christof’s experiments, the skull was opened and a concept implanted in the brain. They were all created by internal processes of the brain. So skull closure is a moot point in this discussion. And “simplicity, automation, and computational efficiency” are not inconsistent with the system being closed and emergent.

Asim



On 7/13/11 5:10 PM, Juyang Weng wrote:
? In none of Christof’s experiments, the skull was opened and a concept implanted in the brain.

Let me be clearer: The proposers of extra-body-concept cells or extra-body-concept grandmother cells must explain how a particular cell in the closed skull can get a particular extra-body-concept while there are many other objects in the background at any time.

When there is no 5-chunk model about the brain-mind, any reasonable hypothesis is fine, including extra-body-concept cells. This is a process for theories to be developed, verified and improved. After the 5-chunk DN model has been proposed, I hope that any new theory is compared with it in terms of 5 chunks:
(1) Development: Can it do skull-closed autonomous development, task-nonspecific, emerging representations?
(2) Architecture: Does it have an biologically plausible emerging architecture?
(3) Area: Does it have emerging area as a building block? Does the area handle both working memory and long term memory?
(4) Space: Can it autonomously learn and attend objects in complex backgrounds while the skull is always closed?
(5) Time: Can it deal with general temporal context and can recombine objects on the fly while the skull is always closed?

We should also check whether old theories are advantageous with regard to the new DN model.

A major problem that I clearly see is that a person needs knowledge in 6 disciplines to understand and appreciate the 5-chunk brain-mind model. This is probably one of the reasons why we have to spend so much time back-and-force through many emails for a minor issue whose answer is very clear from the DN model. Any of the above 5 chunks seems more important to the extra-body-concept cell issue. We should not address the brain-mind problem in a piece-meal way.

-John


On 7/13/11 9:45 PM, Peter Foldiak wrote:
An important, but so far neglected side of this discussion is what we
mean by "thing". When we see a picture of Jennifer Aniston, we seeing
all of the following things, each of which can be considered a "thing":

1) Jennifer Aniston
2) a picture of Jennifer Aniston
3) a photo taken of Jennifer Aniston
4) one particular photo of Jennifer Aniston, presented at a certain size
5) a photo taken of Jennifer Aniston facing to the left, in a blue
dress, alone, lit from above
6) Jennifer Aniston 3 years ago
7) Jennifer Aniston smiling, in the company of other people
8) a long-haired young woman smiling in a way similar to the way
Jennifer Anison smiles
9) a woman
10) a human
11) a mammal
12) a picture of an actress from a TV series
13) an object with long hair-like texture on the side of the central blob
14) one of the images from a random collection of images

There might also be other related "things":
15) a doll made to look like Jennifer Aniston
16) the voice of Jennifer Aniston
etc.

Whichever hypothesis the data supports, you can call that the
"concept-cell" or "grandmother-cell" for that hypothesis. Both the too
specific and the too general hypotheses can be ruled out by control
stimuli. All hypotheses except No. 14 (random selection) are consistent
with the "explicit coding" hypothesis I mentioned earlier.
Some of these are compatible with the data (and can be experimentally
controlled), others may not be.



On 7/14/11 12:04 AM, Ali Minai wrote:
I think that Peter makes a crucial point here. The notion of a "grandmother cell" is ill-defined because there are infinite ways to define "grandmother". When we talk of "concept cells", we are implicitly saying that cell activity corresponponds exclusively to one of the things that we might recognize as a "concept" at the mental level. Who knows what is a "concept" for a cell or a brain region? The concepts we think in terms of are labels we use at the mental/cognitive/linguistic level, and do not need to correspond isomorphically to physical phenomena such as cell activity. Whether they do or not will, in the end, be shown by the experiments being done now with implanted electrodes where experimenters try to decode specific "meanings" (including motor meanings, i.e., actions) from brain activity. The experiments by Tanaka, Yamane and others with monkeys (I think John also cited them) suggest that what will be found is a complex, semi-hierarchical, semi-local but ultimately not completely local code. Whether this involves neurons whose activity is highly specific to one "concept" wil remain impossible to check as Christof points out. We can make theoretical arguments based on coding efficiency, and Peter and Christof are the experts for that.

My hunch (and something we use presumptively in our models) is that the tuning of cells to "concepts" has the following general pattern:

1. In any local neighborhood of, say, IT, there will be a high diversity of tuning in representational space with some local smoothness, i.e., cells that are very close together will tend to be tuned to similar "things" but within a small neighborhood, it will be possible to find cells tuned to almost every "thing" within the domain of specialization of that region (e.g., faces or tools).

2. Over the region as a whole, any given "thing" will produce activity in several locations.

3. The specificity of tuning of cell activity to "things" we recognize cognitively as "things" (concepts) will be extremely variable, and whatever "thing" we go looking for, we'll find cells tuned to it. For example, we'll find "Jennifer Anniston cells", "pretty woman cells", "woman cells", "human cells", etc.). If we looked long enough, we will also find that the many (all?) cells are tuned to very different "things", but "long enough" may be longer than the life of the experiment, the subject, or even the experimenter:-).

I "grew up" on hippocampal place cells, and what I'm suggesting is that cells in IT, PFC, etc., have "place fields" in cognitive space - call them "concept fields" - and will have fields of different widths as well as multiple modes. In this situation, there may well be some cells which, in the limited life-experience of an animal, may only ever express one concept field, and would be "grandmother cells" under our definition, but this does not mean that coding with grandmother cells is the default method being used. In any sparse code, one can find cells for which the tuning appears uniquely specific over any finite set of probes (and that is what a finite life, and certainly a lab experiment, is). Also, the width of tuning would vary a lot even for such cells.

Of course, this view is not original with me and is implicit or explicit in the work of many others. I just find this a useful way to think.

In all this discussion of representation and coding, it is also critical to keep in mind that all this is happening within a physical, dynamical system, and all the phenomena we are discussing must emerge from this dynamics at various temporal and spatial scales. Using that as a constraint can also help answer these questions.

Ali



On 7/14/11 2:05 AM, John Collins wrote:
At various points in this discussion there has been quoted the estimate by Waydo et al (J. Neurosci. 26, 10232 (2006)) that each cell in the MTL responds to 50-150 distinct objects. (They actually say "distinct representations", but I cannot make sense of the sentence without changing "distinct representations" to "distinct objects".)

The estimate is obtained by multiplying their measured sparsity (0.54%) by Biederman's estimate (1987) that a typical human adult recognizes between 10000 and 30000 discrete objects. I think it is worth examining the basis of the estimates in more detail.

Now an essential assumption in Waydo et al's calculation of sparsity from the data was that all the relevant cells have the same sparsity. However, in work with Dezhe Jin (http://arxiv.org/abs/q-bio.NC/0603014) I have found that a much better fit to the data (from Quian Quiroga et al, 2005) is obtained by allowing for two populations of cells, with different values of sparsity. By sparsity, we mean the fraction of images to which an above-threshold response is obtained, in the same sense as chosen by Quian Quiroga et al, and it is taken for granted that as in the data the images all correspond to different objects.

We find that using a single population with one sparsity gives a very bad fit to the data. (Our statistical methods for this are more sensitive than those of Waydo et al.)

As mentioned by Waydo et al, there is an issue of how many cells are silent, i.e., how many give no detectable spikes during the experiment. A first analysis assumes there are no silent cells. In that case we find that the two populations have sparsities of about 2x10^{-4} and 0.04. Call these the ultra-sparse and regular populations. Our fit to the data shows that cells in the ultra-sparse population are 20 times as numerous as those in the regular population. (But their very low sparsity implies that relatively few of them actually give a measured above-threshold response.) With Biederman's numbers, each cell of the ultra-sparse population responds to 2 to 6 objects, much less than the previous estimate. The sparsity fitted by Waydo et al is, to a useful approximation, a weighted average of our two sparsities, the weights being the relative numbers of cells in the two populations. Cells in the regular population respond to many more objects, of course, than those in the ultra-sparse population.

Next we allow for silent cells, as mentioned in the penultimate paragraph of the Waydo et al paper; they could outnumber the detected cells by as much as a factor of 30; call this the silent-cell factor. (I don't know whether Christof would still accept this possibility.) The fit for the sparsity of the regular population is essentially unchanged. But the sparsity of the ultra-sparse population goes down by the silent-cell factor, e.g., to around 10^{-5} for a silent-cell factor of 30. With Biederman's estimate for the number of distinct recognized objects, each of the ultra-sparse cells responds to less than one object.

I recently looked at more recent measurements by Mormann et al (J. Neurosci. 28, 8865 (2008)). In their Fig. 3, they show the numbers of cells in different regions that give particular numbers of responses to the ~100 stimuli used. Three regions show a substantial excess in the one-response bin compared with the expectation from a single-sparsity model. With the two-population model the sparsities are in line with what I fitted to the earlier data; I think this counts as a successful prediction.

The interpretation of our results could use a lot more discussion. For example, one should question the basis of Biederman's numbers; I suspect they seriously undercount the relevant number of memories. But our results suggest that excluding the grandmother cell hypothesis (or the pure concept cell hypothesis) on experimental grounds is harder than often supposed.

John Collins



On 7/14/11 2:17 AM, Asim Roy wrote:
Would anyone argue for or against John Weng's position that we can't
have concept cells for extra-body concepts but only for intra-body ones?
I think it's an interesting argument based on some experiments. My
theoretical argument would be that if the brain is capable of such
abstractions and simplifications, then it should hold for all concepts,
intra or extra. But that's a very general argument and can be wrong.
There indeed can be exceptions.

Anyone willing to venture some arguments for or against extra-body
concepts? What Peter and Ali and others are referring to sound like
extra-body concepts. And Chistof's experiments find extra-body concepts,
however general ("basketball players", a category concept) or narrow
("Jennifer Aniston," a grandmother-like cell??) they are.

Asim



On 7/14/11 2:26 AM, Harry Erwin wrote:
I've been thinking about this argument in the context of my own research. Let me summarise what I was planning to say at the workshop.

I study and model the inferior colliculus (IC), which is at the top of the auditory brainstem. There are four basic types of neurones known there (see Sivaramakrishnan and Oliver, 2001)--onset cells, sustained regular cells, pause-build cells, and rebound cells.

Onset cells respond with a single sodium spike at the beginning of depolarising current injection and sustained regular cells spike periodically during current injection. They can both be modelled in the Hodgkin-Huxley formalism, using sodium and non-inactivating potassium channels. The category that the cell belongs to seems to be based on density of their potassium channels, which open after the initial sodium spike. Onset cells seem to have a higher density of these channels.

Pause-build cells seem to have both non-inactivating and inactivating potassium channels, so that they function like onset cells initially and later shift to sustained regular dynamics. That delay is sensitive to the amount of current injection and the degree to which the cell was hyperpolarised prior to current injection, which means these cells can generate a tuneable delay.

The fourth type of cell--the rebound cell--appears to have T-type calcium channels, which are inactivated at rest. If the cell is hyperpolarised, the calcium channels deinactivate gradually. When the hyperpolarisation is terminated, there is an immediate rebound sodium spike followed by a slow calcium 'hump' or 'spike'. During the hump, depolarising current injection produces a burst of sodium spikes.

Rebound cells seem to have an important role during bat echolocation--they spike in response to echoes during fairly narrow time-frames. They might also play a role in match-mismatch processing, if primed by inhibitory reafference to detect expected patterns of sensory afference during narrow time-frames. Finally, they might serve as amplifiers, converting one or a few incoming spikes into a broad burst of spikes.

We've noticed that these four cell types might work together in speech understanding. If you're aware of the work in the visual system on the letterbox region (Stanislas Dehaene, Reading in the Brain), you are aware that cells in that region seem to respond to pairs of letters separated by 0-2 other letters, so that the next region can then assemble words in parallel and reliably from the component pairs of letters. So those cells might be seen to function as 'concept cells.' I work a lot with dyslectic students, and I understand one of the current models of dyslexia is that it involves problems with speech understanding. A simple network can be created with the cells I've described to detect pairs of sounds separated by an interval, with the signal being a rebound cell producing a burst of spikes. I call these rebound cells 'diphone detectors' or 'diphone cells'. A1 (or perhaps some other auditory region that the IC projects to) might then recognise spoken words directly based on the pattern of activation and without the need to assemble the word incrementally.

Note the IC is believed to be pre-conscious. This process shows how neurones at the preconscious level can gain cognitive meaning.



On 7/14/11 2:28 AM, John Collins wrote:
Asim Roy wrote:


John Collins has an insightful note on concepts cells. I have requested him to post it on this list.

Here it is:

To avoid many of the misunderstandings that plague this subject, I think there are several related ideas that are not always cleanly separated and that need to be defined sufficiently precisely. (I will take for granted that the definitions may need adjustment later to correspond to reality, etc.) To my mind there are four related ideas that need to be defined. Here's my attempt.

1. First we need the idea of a concept representation. This is a pattern of neural firing that codes that a particular concept is active, e.g., as recognition of a certain entity from a stimulus. I would want to use the term "concept" in the broadest possible sense.

2. A concept representation is to be contrasted with what I call a feature representation, i.e., a coding of the components of a particular entity in a stimulus. The distinction between concept and feature representations is not normally made explicit in the literature. One need for concept representations is exactly the efficiency argument Asim mentioned. (Efficiency matters in biological systems, e.g., for good reaction times, for improved speed of learning from experience, and for relatively low energy consumption. So there should be evolutionary pressure for efficiency in neural systems.)

The next question is what kind of a neural representation is used for concepts.

3. A concept cell is one that only fires above some threshold when the concept is active. It does not fire above threshold for any other concepts. For a given concept, there may be more than one cell. Note that multiple concepts may be active simultaneously. This is the appropriate generalization of the original idea of a grandmother cell, in my opinion.

I allow that sub-threshold firing may occur, and the obvious interpretation is as some kind of partial match between a stimulus and a concept. (As in Asim's example of word neurons.) Threshold is to be understood as something like the criterion used by Quian Quiroga et al (2005) for a responsive cell.

4. A distributed concept representation is one where one can decode the concept unambiguously only from the firing of multiple cells. I think an implementation with considerable algorithmic advantages is something like Marr's coden representation: Here a concept is represented by the (above-threshold) firing of k cells chosen at random from a population of N cells, with k being very much less than N.

In practice it can be difficult to distinguish a pure concept cell and a cell in an ultra-sparse concept representation. It may be useful to have two terms, e.g., "pure concept cell", "ultra-sparse concept cell".

As several people correctly say, concept representations (especially pure concept cells) must arise during the lifetime of an organism in response to experience. Concept cells for all possible concepts cannot be present at birth. This statement by itself disposes of many of the a priori arguments against grandmother cells. However, pure concept cells are most simply implemented if one has adult neurogenesis. In mammals, adult neurogenesis appears to occur only in the dentate gyrus of the hippocampus and in the olfactory bulb. This argues against the concept cell idea anywhere but in the hippocampus. (However, there may be other ways of achieving what I would call a concept node without actual concept cells. Another long story here...)

The situation regarding adult neurogenesis is different in certain other species, so statements about the existence or non-existence of pure concept cells need not be universal across species.

A final remark is that if a certain cell gives an above threshold response to a certain concept that does not by itself imply that the cell is part of the representation for that particular concept. For example, a hippocampal cell might code a particular episodic memory, and would respond to stimuli containing particularly important parts of the memory. I think this is very plausible; Koch and his friends advocate something like this view; it explains quite a lot of the properties of their data.

If you think of an episode as a kind of generalized concept, then this idea does not even need a change in the idea of a concept cell. I envision that in response to a stimulus, multiple episodic memories are (re)activated.

John Collins


On 7/14/11 2:45 AM, Yoonsuck Choe wrote:

Asim,

Here's some thought.

Thinking about simpler and more fundamental concepts can be more productive. For example, think about the concept of "space" and the dimensionality of the space occupied by an agent. Given no other hand-crafted external information, can an agent with a set of simple sensors and actuators figure out the dimensionality of the space it occupies?

Philiopona, O'Regan, and Nadal showed that it is possible. The significance of this result is that based on purely "internal" information, properties of the "external" world can be inferred. I think this could be a counter example to John's agrument that there cannot be concept cells for extra-body concepts (in a very broad sense), although in Philipona et al.'s work there is no explicit mapping of this knowledge to a single cell.

@Article{philipona:nc03,
author = "D. Philipona and J. K. O'Regan and J.-P. Nadal",
title = "Is There Something Out There? {I}nferring Space from
Sensorimotor Dependencies",
journal = "Neural Computation",
year = 2003,
volume = "15",
pages = "2029--2050",
url = "http://nivea.psycho.univ-paris5.fr/Philipona/space.pdf",
bibdate = "Fri Feb 28 16:49:38 CST 2003",
bibauthor = "yschoe",
annote = "
",
}

Yoonsuck
choe@tamu.edu



On 7/14/11 3:52 AM, Jay McClelland wrote:

I fully agree with Peter on this. In fact, in an earlier post in this discussion, I raised this issue, and it was discussed at length in Plaut and McClelland's reply to Bowers grandmother article. This issue was also discussed in McClelland & Rumelhart (1985) as one basis for preferring the use of distributed vs localist representations. That paper was written in response to a version of the current controversy that arose strictly within a psychological context, ie within the context of exemplar theory. Rumelhart and I asked just what the exemplars were supposed to be exemplars of, ranging from general concepts to specific exemplars to particular experiences with particular exemplars etc, just as Peter does below. Of course Rosch's work on the basic level provides one kind of answer to this question, but this was never fully satisfactory. I still believe distributed representations, subject to differentiation with experience, and sensitive to patterns of coherent covariation, provide the right way to think about our representations of objects. They exhibit tendencies toward discreteness but are not inherently discrete, solving many of the classical problems that have always faced theories of conceptual representation. These issues were also discussed in an AI representation context even in the 70's. E.g. how many different types of restaurant Scripts was it necessary to posit an individual had, under Shank's script theory of restaurant knowledge? Did you need a different script for fancy restaurants, fast food restaurants, drive-thru restaurants.... and a special one for that famous seafood restaurant in Boston called the non-name restaurant? Rumelhart and I both felt distributed representations provided a better way of thinking about this issue.

I will try to read more on this thread and comment further when I get a moment.

Best to all,

-- Jay McClelland



On 7/14/11 8:04 AM, Asim Roy wrote:
I think I have become the de facto promoter of the findings by Christof and his associates. But I am glad they are doing these experiments and I believe Christof is writing a book about all this. Can’t wait to get a copy of it. But wanted to draw your attention to their 2009 paper on discovery of triple invariant cells. That’s a very high level of abstraction encoded in a single cell. In the following, I will summarize their study and also use some direct quotes from the following paper:

Quian Quiroga, R., Kraskov, A., Koch, C., & Fried, I. (2009). Explicit Encoding of Multimodal Percepts by Single Neurons in the Human Brain. Current Biology, 19, 1308– 1313.

Quian Quiroga, Kraskov, Koch and Fried (2009) report they found that single MTL neurons can encode information about the same percept that can arise in different modalities such as visual, textual and sound. For this experiment, they implanted 7 subjects with microelectrodes and recorded from 750 MTL units (335 single units and 415 multiunits; 46.9 units per session) over 16 experimental sessions. Of the 750 units, only 79 had any significant response to at least one stimulus. For the neurons that responded, they checked their modality invariance properties by showing the subjects three different pictures of the particular individual or object that a unit responded to and to their spoken and written names. In these experiments, they found “a neuron in the left anterior hippocampus that fired selectively to three pictures of the television host Oprah Winfrey and to her written (stimulus 56) and spoken (stimulus 73) name…. To a lesser degree, the neuron also fired to the actress Whoopi Goldberg.” They also found a neuron in the entorhinal cortex of a subject that responded “selectively to pictures of Saddam Hussein as well as to the text ‘Saddam Hussein’ and his name pronounced by the computer….. There were no responses to other pictures, texts, or sounds.”

A most interesting finding is about the researchers who conducted these experiments and how they were quickly encoded as a percept in the MTL. They found a neuron in the amygdala that was “selectively activated by photos, text, and sound presentations of one of the researchers performing recordings with the patient at UCLA.…..Altogether, we found five units responding to one or more researchers performing experiments at UCLA….None of these researchers were previously known to the patient, thus indicating that MTL neurons can form invariant responses and dynamic associations—linking different individuals into the same category ‘the researchers at UCLA’—within a day or so.” The authors call these neurons “triple invariant” neurons and they were those that had the visual invariance property and also had significant responses to spoken and written names of the same person or object. They found 17 of the 79 responsive units to have such triple invariance property. They report that “Eleven of the neurons showing triple invariance responded to only one person and the remaining six responded to more than one person or object.” They conclude that these findings show that information from different sensory modalities converges onto neurons in the MTL.

Thus, some triple invariant neurons seem to encode a very high-level abstraction about a single person – from picture to voice to written name. That indeed is very revealing.

I wonder if such a neuron can provide the linkage between say a name and the persons voice and image. If a single triple invariant neuron does indeed provide such a mapping, that could change our thinking about how semantic coding is done in the brain.

Asim



On 7/14/11 9:29 AM, Juyang Weng wrote:
Peter:

What you said is called concept learning and concept abstraction by a brain-like network. You did not mention backgrounds and attention, which are necessary for extracting concepts from a vast cluttered scene.

These fundamental issues must be addressed by the DN brain-mind network. Theoretically done, in a general purpose way, with some experimental supports in WWN-1 through WWN-5. WWN-1 through WWN-4 dealt with two concepts, type and location. WNN-5 added a third concept: scale. In a paper that is under the third round of revision, I listed 22 concepts.

-John



On 7/14/11 9:50 AM, Juyang Weng wrote:
Asim,

Of primary importance is not whether there are concept cells for extra-body concepts inside the closed skull. This issue is secondary.

Of primary importance is that
(1) without any concept cell for extra-body concepts
(2) there is already a brain-mind model (DN) in the 5-chunk scale
(3) that does concept learning and abstraction in cluttered scenes
(4) perfectly (error-free) for training data if there is sufficient resource (brain size) and
(5) optimally for a limited resource (brain size and learning time).
I will present it during our session in IJCNN 2011, San Jose.

I am not saying the DN model has shown all the brain functions, as any model of nature is an approximation.

-John



On 7/14/11 10:04 AM, Juyang Weng wrote:
Harry,

I am sorry if our views are different, but it is good to communicate so that we can understand each other.

You are modeling functions for extra-body concepts of neuronal cells. This is something that traditional AI got stuck for over 60 years.

In my view,
(1) all brain cells must emerge, must not be rigidly specified by the genome (i.e., not handcrafted as you did)
(2) the genome program (DP) regulates to a great deal the functions of cells for intra-body concepts, such as arm, muscles, skull.
(3) the genome program (DP) regulates inborn behaviors to a great deal since they are intra-body concepts.
(4) the genome program (DP) only regulates the learning mechanisms for extra-body concepts. For your "speech understanding" model, I suggest that you model the genome program (DP), not the phenotype from the DP. What you did is called GOFAI, handcrafted, symbolic, static, not emergent, and brittle. The intelligence of the system is in YOU as the central controller, not the system.

-John



On 7/14/11 10:23 AM, Juyang Weng wrote:
Yoonsuch:

The model of Philiopona, O'Regan, and Nadal 2003 is not a fully emergent model, not skull closed. Its result cannot say much about the skull-closed emergent brain.

Read: "We will now search for the relationships between the dimensions of all the entities we have defined." This assumes that the baby has an inborn concept
about the external environment: "Dimension". However, a normal child does not learn this concept till late in a school. An extra-body concept like "dimension" must be emergent, like those in DN.

-John



On 7/14/11 10:50 AM, Harry Erwin wrote:
In my research work on bat target capture behavior, I came to a conviction--influenced to be sure by Don Griffin--that the bat lives *in* an internal model of the external world, which is updated asynchronously and usually at a low rate from sensory afference. I'm not sure that can emerge without some genomic programming.

On the other hand, I wasn't suggesting the diphone networks were rigidly specified by the genome. In fact, I suspect they are the result of the small child hearing a language being spoken. However, I also suspect diphone networks exist in all mammals.



On 7/14/11 11:15 AM, Juyang Weng wrote:
In a paper under the last round of review by TAMD, I cited Koch & coworker's work on multimodal invariant cells, as our WWN model gives a computational account for the emergence of such cells. However, as I explained in my earlier emails, in WWN there is no cell that is a concept cell for any extra-body concept in a rigorous "if and only if" sense, except the case wherein we teach each motor error to represent a single value of an extra-body concept (which we did!). This is highly unlikely for a human, as each concept (e.g., saying "Saddam Hussein") requires many motor cells to work together to express.

-John



On 7/14/11 12:18 PM, Asim Roy wrote:
John, how would you characterize the triple invariant neurons. Would
they qualify as pure concept cells?

And you mentioned neurogenesis being necessary for concept cells, but
that there could be a way around it. Without a very long story, I am
curious how that could happen if neurogenesis occurs only in the dentate
gyrus of the hippocampus.

Asim




On 7/14/11 3:54 PM, Asim Roy wrote:
I certainly agree with Jay that distributed representation has certain
advantages. However, given the mounting evidence for concept cells,
triple invariant neurons and so on, it is not easy to discard or
overlook the idea that there could also exist very high-level
abstractions and encodings in the brain. As John Collins pointed out,
there is definitely an efficiency aspect to these abstractions,
particularly for repeating patterns. My view is that the brain uses both
localist (in the form of concept cells, triple invariant neurons and so
on) and distributed representation. I view localist representation as
the limiting case of distributed representation. Thus there is no
inherent conflict between the two, at least in the mathematical or
theoretical sense. It's not really one or the other. It's both. Perhaps
both are needed for efficient operation of the brain. Much of our
argument is about localist vs. distributed representation. It need not
be that way.

Asim



On 7/14/11 5:05 PM, jeff bowers wrote:

I agree with Peter that it is important to be clear what we mean by “thing”. I think some things in his list need to be treated differently than others. Here is his list.

1) Jennifer Aniston
2) a picture of Jennifer Aniston
3) a photo taken of Jennifer Aniston
4) one particular photo of Jennifer Aniston, presented at a certain size
5) a photo taken of Jennifer Aniston facing to the left, in a blue
dress, alone, lit from above
6) Jennifer Aniston 3 years ago
7) Jennifer Aniston smiling, in the company of other people
8) a long-haired young woman smiling in a way similar to the way
Jennifer Anison smiles
9) a woman
10) a human
11) a mammal
12) a picture of an actress from a TV series
13) an object with long hair-like texture on the side of the central blob
14) one of the images from a random collection of images

There might also be other related "things":
15) a doll made to look like Jennifer Aniston
16) the voice of Jennifer Aniston
etc.
The first issue to note is whether we are talking about an episodic memory of Jennifer Aniston (a cell in the hippocampus), the concept/semantic representation of Jennifer Aniston (perhaps in the anterior temporal lobes) or perceptual representation of her face (perhaps in IT). I think grandmother cells were traditionally related to perception, but the same issue relates to all domains.
Let’s just consider perceptual representations of faces. On my understanding, a grandmother cell theory would predict that any recognizable image of Jennifer Aniston (either due to seeing her in person, looking at a photo, whether the photo was taken yesterday or from 10 years ago, facing left/right, etc.) would activate a grandmother cell of Jennifer Aniston in IT (there would presumably be many). Grandmother cells code for perceptual category, and all these images are members of a fixed category.
Now of course, Jennifer Aniston looks different in 1-7, so the brain can’t be coding for Jennifer Aniston in the same way. But a grandmother cell is not thought to encode all the perceptual details of Jennifer Aniston – rather it is a cell that categorizes all the different instances of Jennifer Aniston as Jennifer Aniston. The main task of object recognition is to identify an object in different contexts (regardless of orientation, size, color, etc), and a grandmother cell theory is simply a claim that the invariance is achieved by activating the same cell in all the different contexts (a cell at the top of a visual hierarchy). Of course, when you see Jennifer Aniston from the left and the right, up close, or at a distance, a very different set of simple cells, complex cells, etc. etc. will be activated in the hierarchy, and this will lead to different visual experiences. But we recognize all the different visual experiences as Jenifer Aniston by virtue of a Jennifer Aniston cell be activated in all cases.
Different cells would be needed for 8-10, as Jennifer Aniston is not being seen. There would not be a “mammal” cell, as this is a conceptual category, not a perceptual category. Neither would there be one for Jennifer Aniston’s voice (IT is a visual area).
A key claim of a grandmother cell is that representations of objects is “context independent” – that is, the equivalence of Jennifer Aniston in different contexts is coded by the same invariant representation (the same grandmother cell is activate in 1-7). This is the key idea of a “symbolic” representation – a representation that is invariant in different contexts. A key claim of distributed representations is that representations are not context independent. On this view, different views of Jennifer Aniston produce similar patterns of activation, but there is no invariant representation. John Hummel makes a good case for the computational importance of symbolic representations, both in the case of object perception and thinking. Of course this is all controversial, but it is how I understand the issues.
jeff



On 7/14/11 10:05 PM, John Collins wrote:
Asim Roy wrote:
John, how would you characterize the triple invariant neurons. Would
they qualify as pure concept cells?

Not necessarily. A pure concept cell is one that fires (above threshold) for just one concept (whatever the concept may be) and not for any other concepts. The triple invariant neurons qualify as candidates for pure concept cells. But it is also possible that concepts are represented by abstract ultra sparse code words (to use Peter's terminology). In that case the triple invariant neurons could fire above threshold for a small number of other unrelated concepts. Within the range of stimuli used for the experiment, these neurons behave as if they are pure concept cells, and I think give clear evidence that high level concepts are being coded neurally.

A direct test would need too many stimuli to be practical, as Christof says. But I think statistical analyses improving the one Jin and I did can do better. The question could be answered much better if one had a sufficiently complete understanding of how, for example, the episodic memory system works in the hippocampus.

And you mentioned neurogenesis being necessary for concept cells, but
that there could be a way around it. Without a very long story, I am
curious how that could happen if neurogenesis occurs only in the dentate
gyrus of the hippocampus.

Analogs of concept cells could occur at the subneuron level. They could be particular branches of a dendritic tree. The structure of the dendritic tree can change with experience without a change in neuron number. A rather clear example of such changes is in the mushroom bodies of honey bees -- Farris, Robinson and Fahrbach, J. Neurosci. 277, 3087 (2010).

John Collins



On 7/15/11 2:01 PM, Juyang Weng wrote:
> A key claim of a grandmother cell is that representations of objects is “context independent”

Jeff: This is inconsistent with the fact that the skull is closed from conception. If your theory has a biological basis, explain how this grandmother cell develops using knowledge in biology. Talking about neurons does not mean that your theory has sufficient supports from known facts in neuroscience and biology.

Asim: This is also inconsistent with your position about autonomous learning. How does your system learn autonomously if you require such grandmother cells (or concept cells) in your system? You have to resort to a human to "intervene" as the central controller of the brain. All symbolic representations require a human designer to hand-pick a set of symbols (like the one above) but the meanings of those symbols are only in his mind. After he talks to other human programmers about his symbols using a human language, each of the other programmers has similar, but slightly different, meanings in his mind. The example from Peter is good to consider such a diversity. The meaning of a symbolic concept (such as Jennifer Aniston) is consensual, as I argued in my paper "Symbolic Models and Emergent Models: A Review" to appear in the next issue of TAMD.

-John



On 7/15/11 2:14 PM, Ali Minai wrote:
John

The point you make (that all symbols/concepts are necessarily subjective and therefore non-universal) is a valid one, but I think that using the "skull being closed at birth" complicates this argument unnecessarily. The point is what you say later in your post: That each mind (with all its symbols/concepts) must emerge from each individual brain, and does not draw upon some Platonic universal concepts. I don't think that this precludes the emergence of cells that are effectively grandmother cells, but not because someone tuned them to an objective concept. It is just that they fire very rarely and for highly similar things. That doesn't make grandmother cells the foundational principle of mental representations.

Best

Ali



On 7/15/11 3:39 PM, Asim Roy wrote:
I think Christof describes emergence of concept cells quite well. And there is no outside intervention.

“Every time you encounter a particular person or object, a similar pattern of spiking neurons is generated in higher-order cortical regions. The networks in the medial temporal lobe recognize such repeating patterns and dedicate specific neurons to them. You have concept neurons that encode family members, pets, friends, co-workers, the politicians you watch on TV, your laptop, that painting you adore….Conversely, you do not have concept cells for things you rarely encounter, such as the barista who just handed you a nonfat chai latte tea.”

As far as I understand, none of the theories of representation are denying the existence of “meaning” in the brain in some physical form, whether its subjective, objective or evolves over time. In distributed representation, symbols and meanings are encoded in patterns of activations. It seems like a class of patterns, if it’s repeating (e.g. repeating images of Jennifer Aniston), is simply handled more efficiently by dedicating a neuron to it. In some sense, you have “meaning” (a basic interpretation of the stimulus) in both forms, localist and distributed. The localist ones (concept cells, invariant neurons) actually owe their existence to the distributed form. You can think of it as just an additional encoding of the meaning of a class of stimuli. The distributed representation has not disappeared, it’s still there.

Again, I believe none of the theories deny the existence of “meaning” in some physical form. And there can be other kinds of meaning beyond a basic recognition of a stimulus. And they all have a physical existence, whether its subjective, objective or evolves over time. As far as I know, neuroscience got rid of the concept of “soul” a long time ago.

All, please feel free to correct me on these statements, particularly if any of the interpretations are wrong.

Asim



On 7/16/11 1:19 PM, Walter J Freeman wrote:
I write to rescind my endorsement of 'concept cell' in view of its multiple ambiguities. I suggest an alternative: 'Hebbian cell', i.e., a member of a nerve cell assembly as first formulated by Donald Hebb in 1949.

I believe the Hebbian cell first evolved in the olfactory bulb, which is a type of allocortex. Only a few categories of odors are important for most animals (types of food, mate, predator). Each is defined only by the selection of chemoreceptors as examples. Each sniff coactivates a few of the available sensitive receptors, which coactivate a subset of mitral cells. Strengthening of the synapses interconnecting the subset on serial sniffs defines an assembly with respect to the category of the stimulus for the subject. It is updated with each use. A stimulus to any subset ignites the whole assembly. The operations performed are inductive generalization, abstraction by deletion of information about which equivalent receptors were activated, amplification, and broadcasting. Partial overlap with other assemblies is not excluded; it may promote associations.

Evolution derived neocortex from allocortex. The four basic operations have been elaborated in series preceding cognition. I conclude that odor cells, place cells, mirror neurons, Jennifer Aniston cells, etc. are Hebbian cells. Each assembly generalizes, abstracts, amplifies, and broadcasts a signal that selects a mesoscopic attractor from a landscape constituting memory. Importantly, it provides the transition energy that is required for a phase transition by which cortex recollects a memory. A microscopic sensory input triggers a macroscopic perception, which is the door to cognition. The number of Hebbian cells in each assembly (10^4-10^5) suffices for detection by random sampling. The cells fire a brief, vigorous burst, but only following onset of a meaningful trigger.

I think that the most compelling argument for this hypothesis is that the Hebbian synapse strongly promotes oscillations in the beta and gamma frequency ranges by facilitating negative feedback. Simulations show that a 20% increase in synaptic gain with reinforcement learning causes >50-fold increase in the rms amplitude of a gamma burst.
On 7/16/11 1:27 PM, Ali Minai wrote:
Dear Walter

Thank you for injecting the notion of evolution into the discussion. I think it is extremely important and will help to resolve important issues. The scenario you suggest is very plausible and illuminating.

Ali



On 7/16/11 3:41 PM, Juyang Weng wrote:
Walter:

Your email is certainly a move toward a more productive direction, in my personal view.

> The number of Hebbian cells in each assembly (104-105) suffices for detection by random sampling.
Hebbian mechanisms, like STDP, are indeed very powerful.

> suffices for detection by random sampling
Random sampling of what, sensory inputs?

However, Hebbian mechanisms are not sufficient to characterize why each neuron highly selectively connects a relatively small number of neurons among many in each area.

Synapse maintenance, modeled (first?) in Wang, Wu & Weng, 2011 (IJCNN), while working with Hebbian learning mechanisms, seems to play a more important role in enabling "highly selective connections" among many possible neurons than Hebbian mechanisms alone. It seems to explain why only about 1/5 of neighboring neurons connect with each other, as reported by Andrew Callaway. Synapse maintenance determines precisely which neurons to connect among many neurons in an area, while Hebbian mechanisms determine the efficacy of each connection.

Prior network models assumed designed connections. We have found such statically designed connections are not sufficient, e.g., for segmenting a stable object view along its precise contours from cluttered backgrounds. As far as I know, synapse maintenance in Wang, Wu & Weng, 2011 (IJCNN) seems to be the first model to address this hardly-ever-addressed tough problem.

-John



On 7/16/11 4:26 PM, Juyang Weng wrote:
Ali,

Why does the skull-closure prevent the emergence of grandmother cells or concept cells? Let me explain in more detail.

Biology:
Each brain is skull-closed. Each neuron is autonomous. It uses its own cellular mechanisms to "swim", change its shape, and connect, in the "sea" of neurons.

GOFAI/Many Cognitive Scientists:
Suppose that the skull is open. Human designer designs a symbolic representation, where each symbol is explained by the design document in a human natural language. The natural language corresponds to the consensual meanings among human communicators. However, each text sentence (e.g., Jennifer Aniston)
of the natural language implies a slightly different set of meanings in each human mind, like the long list that Peter Foldiak supplied in an earlier post. The human designer does not care about that. He only cares about what he meant in his design document. As the skull is incorrectly treated as open, the human designer "implants" his symbolic representation into the skull-open machine "brain". He thought that he did correctly. He argued with connectionists for over 30 years. Marvin Minsky said that such symbolic representations are clean and logic.

Weng:
The brain is skull-closed. The above symbolic representation is only what is meant by the single human designer's actions when he wrote text of natural language in his design document, not necessarily the same to other human communicators. Each neuron inside the skull-closed brain cannot figure out the meanings of the human designer in his design document. It cannot figure out the meanings to another human communicator when he reads the design document. "Grandmother cells" and "concept cells" are special case of such symbolic representations.

I have attached the galley proof of my paper "Symbolic Models and Emergent Models: A Review", which explains more about why any symbolic representation is impossible for a biological brain. If anybody still likes to argue for "grandmother cells" or "concept cells", please at least read this paper first.

-John



On 7/17/11 9:43 PM, Walter J Freeman wrote:
Dear John,

I agree that the Hebbian mechanism is only one early step in the process of cognition, but it is particularly relevant to the question, 'do spiking neurons have meaning?' The answer is, 'no', but in generalizing and abstracting in the light of experience unique to the individual it is the first step from meaningless sensation toward mind.


> The number of Hebbian cells in each assembly (104-105) suffices for detection by random sampling.
Hebbian mechanisms, like STDP, are indeed very powerful.

suffices for detection by random sampling
Random sampling of what, sensory inputs?

Random sampling of neurons by probing with a microelectrode one at a time among 10^5/mm^3 neurons in cortex.


However, Hebbian mechanisms are not sufficient to characterize why each neuron highly selectively connects a relatively small number of neurons among many in each area.

Right, Hebb made no mention of reinforcement, which is essential to prevent habituation and invoke association, which does selectively increase synaptic gain.

However, memory is not in a small number of neurons. Consolidation involves changes in synaptic gain of an indefinite number of neurons. On average each neuron receives from 10^4 neurons and transmits to a different 10^4 neurons, yet connects with <1% of neurons within its dendritic arbor. The probability of reciprocal connection between pairs is 10^-6. Hence apart from I/O networks most cortical neurons interact with their milieu by responding and contributing to pulse densities. The role of the Hebbian assembly is to boost microscopic input into macroscopic dynamics, where the rules of interactions among populations differ from the rules in neural networks.

I think that this difference in levels of analysis may in large part account for the differences between our interpretations of the literature we share.

Walter



On 7/18/11 10:27 AM, Peter Foldiak wrote:
Horace Barlow's "Cell assemblies versus single cells" seems to be relevant here (particularly on the 'imaginary virtues of cell assemblies')

http://www.trin.cam.ac.uk/horacebarlow/157.pdf



On 7/18/11 4:50 PM, jeff bowers wrote:
Interesting paper! I had not considered the discrepancy between simple cell selectivity and psychophysical measures -- such a simple point,


jeff



On 7/19/11 7:01 AM, John Collins wrote:
Juyang Weng wrote:


The brain is skull-closed. ... "Grandmother cells" and
"concept cells" are special case of such symbolic representations.
I have attached the galley proof of my paper "Symbolic Models and Emergent Models: A Review", which explains more about why any symbolic representation is impossible for a biological brain. If anybody still likes to argue for "grandmother cells" or "concept cells", please at least read this paper first.

I agree that all or most of representations in the brain are emergent. (E.g., representations of language obviously depend on an individual's experiences.)

In the paper John sent out, definition 3 is that "A symbolic representation in the brain of an agent contains a number of concept zones where the content in each zone and the boundary between zones are human handcrafted."
What is wrong in principle with eliminating the requirement of human handcrafting? I.e., why shouldn't a symbolic representation be emergent? Of course, the particular pattern of neural firing that correspond to a particular symbol will not have its meaning determined a priori. Instead the semantics would come from the emergent pattern of synaptic connectivity and synaptic strengths. Similarly the particular set of meanings that are represented will also be emergent.

A bit later on, there is a discussion of finite automata, and it seems to be indicated that these have to be static handcrafted affairs. Again what is wrong with an emergent finite automaton? A biological case where there is something remarkably like a finite automaton that is also demonstrably emergent is in the song production system of zebra finches. See R.H.R. Hahnloser, A.A. Kozhevnikov, and M.S. Fee, ``An ultra-sparse code underlies the generation of neural sequences in a songbird'', Nature 419, 65--70 (2002).

On page 13 is written "An emergent representation is harder for humans to understand, as it is emergent in the sense that a particular meaning is not uniquely represented by the status of a single neuron or by a unique firing pattern of a set of neurons." The difficulty of understanding I agree with, but I don't understand why a meaning could not be represented by the status of a single neuron or by a unique firing pattern of a set of neurons. Just as in the case of an emergent symbolic representation, the firing pattern for a given meaning is different between different individuals, as is the set of represented meanings. Thus we cannot have uniqueness between individuals. However, what is impossible with having a particular meaning represented by a particular (emergent) pattern of firing that is unique within a single individual brain?

On p. 19 and 20 is mentioned that with c concepts each taking v values, a symbolic model potentially requires v^c different states, which is obviously impossible for even unrealistically modest values of c and v. But I think this objection is overcome if a symbolic representation (or grandmother-cell representation) is emergent and created only for states that are experienced. It is difficult to imagine more than one such memory being created per second. This amounts to a limit of about 10^9 remembered states per human lifetime. The true number is presumably considerably less.

I don't want to argue here that grandmother cell representations and the like are actually used in the human brain, but only that the arguments given don't seem to me to be sufficient to rule them out a priori, if they are emergent representations.

John



On 7/19/11 7:03 AM, Asher Evans wrote:
Hi All,

My name is Asher Evans. I'm a Ph.D. student at Penn State University working with John Collins on computational neuroscience and theories of knowledge representation in neural networks. I would like to contribute a few comments to the discussion.

I agree that there are many levels of specificity, which Peter Foldiak has pointed out, at which "thingness" may be attributed by a nervous system to a stimulus (and I wish to use "stimulus" in the most general sense possible to include not only sensations but also mental events we are more used to thinking of as internal, e.g., emotions, thoughts and intentions). My suggestion is that there could perhaps be a collection/assembly of concept cells for any such "thing" (concept) at any level in the event that it repeats itself often enough and is relevant enough to the organism's purposes for it to be useful to recognize. I think this is what Cristoph Koch may be getting at in the quotation Asim often brings up.

I don't see it as problematic for the "receptive field" boundaries of these concept cells to be unsharp, cf. fuzzy sets (Zadeh 1965).

I think this quote reflects one of John Weng's main points: "No neuron is in a position to represent any one word, if the skull is closed."

But what if the very reason for the word's existence is its linkage with a neural collection representing a concept in people's brains? And what if the formation of concept neuron receptive fields is highly informed by their attachment to words as used by the people in one's environment?

It does strike me that there is some truth to the notion that neurons can only represent so-called intra-body concepts. However, since sensory neurons are inside the body and are also connected to the environment, this makes it possible for the intra-body concepts to reflect external entities, which seems to be precisely what the identified neurons are doing. This may be related to what Harry Erwin says, "the bat lives *in* an internal model of the external world..." which would obviously apply to humans as well.

I don't know that the skull can altogether be said to be "closed." I think it's at least perforated.

Cheers,
Asher



On 7/20/11 10:11 PM, Juyang Weng wrote:
? the rules of interactions among populations differ from the rules in neural networks.

What do you mean?



On 7/20/11 10:24 PM, Juyang Weng wrote:
> why shouldn't a symbolic representation be emergent?

John, it is important for any theory to have a working model that has acceptable performance from real data.
I guess that this is a reasonable constraint for any theory, since otherwise we would end up arguing back-and-forth without tangible progress.
You can try to answer your question. I already provide reasons that the meanings of any symbol are hand picked.

Symbols are consensual among human communicators. Read that paragraph in the paper carefully and let me know whether it is reasonable. There are details there.

-John



On 7/20/11 10:42 PM, jeff bowers wrote:
Dear John, I'm not sure I understand your point regarding the fact that the brain is skull-closed. What is clear is that neural networks can learn localist representations. Some of the early work on this was reported by Grossberg:

Grossberg, S. (1980). How does a brain build a cognitive code? Psychological Review, 87, 1-51.

Carpenter, G.A., and Grossberg, S. (1987). A massively parallel architecture for a self-organizing neural pattern recognition machine. Computer Vision, Graphics, and Image Processing, 37, 54-115.

Indeed, the learning in ART can be self-organizing, and doesn't require a teacher, as is the case for error driven learning used in back-propagatation. In addition, the learning processes in ART are designed to be biologically plausible (unlike back-propagation). ART might be wrong, but at the very least, it provides an existence proof that localist codes can be learned. There is nothing inconsistent with learning and “winner-take-all” dynamics in a network.

Jeff



On 7/21/11 2:32 AM, Asim Roy wrote:
Unlike constructive algorithms like ART, radial basis function networks
and such, PDP connectionism actually starts with a predefined network,
much like the models of AI. That's been pointed out in several of my
papers in the 90s. Here's the typical reaction you get when you point
out the predefined nature of these connectionist models:

"At any rate, no model can get something from nothing and connectionist
models are no exception. If this is what is meant by "emergent", then
connectionism is not emergent, but neither is the brain or anything
else."

Asim



On 7/21/11 3:23 AM, Christof Koch wrote:
I second Joh and Jeff. We've used totally an unsupervised EM-type learning rule plus sparse coding to expose networks patterned upon Poggio's HMAX to various images.
They cluster these input images such that individual neurons emerge that represent only images of planes, or motorcycles or cars or even specific individuals (e.g. Jennifer Aniston). All we needed was a sparse representation, ie the assumption that the astronomically large number of possible input images are mediated by a small number
of causes. Nothing else is needed. Who knows whether the brain does it like this but this constitutes clear proof that 'concept' cells could emerge in this way without
opening the skull.

Cheers

Christof



On 7/21/11 9:47 PM, Walter J Freeman wrote:
I mean the difference between neurons and a population is like that between the sparse reactant molecules in a chemical reaction embedded in H2O molecules versus a drop of water. For example, neural networks don't form vortices in vector fields the way cortex does. There are biophenomena one cannot 'see' with a microelectrode or 'explain' with Newtonian point processes.

See you in San Jose.



On 7/21/11 10:05 PM, Juyang Weng wrote:
Water, your thought is a natural when we did not understand how individual autonomous cells give rise to body and brain-mind.
My 5-chunk model explains that mechanisms for single cell are sufficient to explain the brain-mind. This is the power of theory, going beyond phenomena from experimental biology. Unfortunately, experimental biologists do not have sufficient background in computer science, electrical engineering and mathematics to understand my IJCNN 2010 paper about the 5-chunk brain-mind. Hopefully, the Brain-Mind Institute at MSU can serve researchers to get over this hurdle.

-John



On 7/21/11 10:20 PM, Juyang Weng wrote:
Water,

thank you for your paper. From my theory, gamma waves and theta waves are emergent phenomena of mutual inhibition and excitation among cells.
They are side effects when the brain tries to sort out the winner neurons. The main point is to sort out the winners, not the side effects.
We contain such side effects using top-k competition among neurons. Such waves can be thought of the fluctuations when the brain "sort".

If you are interested, we can work together to experimentally and theoretically prove/disprove this point. This line of work seems very important.

-John



On 7/21/11 11:22 PM, Juyang Weng wrote:
By brain is skull-closed, I mean that the human programmer cannot program any task specific information into the brain. All symbolic models violate this "skull-closure" requirement.

In terms of ART, this following was in my TAMD review paper:

"ARTMAP \cite{Carpenter91} is a nearest neighbor like classifier for each bottom-up vector input in which each component is a feature value (symbolic, not natural image). After memorizing all sufficiently different ascending input vectors as prototypes, it uses descending signals to successively suppress the current prototype under consideration in order for the next prototype to be examined for the degree of match with the input. Thus, the top-down signal is not part of the features in ARTMAP, but as attention signal in the consecutive nearest neighbor search."

I do not know how many people truly understand ART, although it has been around for many years. ART is a network way to implement a nearest classifier in pattern recognition. Each input must belong to a single object. Each object must belong to one of the known symbolic classes. In contrast, the brain does not require that each image belongs to one of the known symbolic classes.

ART is a SYMBOLIC model, as Stephen Grossberg admitted explicitly in one of his papers. I quote from Carpenter, Grossberg and Reynold NN, 1991: "This map does not directly associate exemplars a and b, but rather associates the compressed and symbolic representations of families of exemplars a and b."

The exemplar for ART cannot be an image which contains multiple objects in a cluttered scene, which WWN deals with. The skull-closed brain must deals with
multiple objects in a cluttered scene.

-John



On 7/21/11 11:47 PM, Juyang Weng wrote:
> Unlike constructive algorithms like ART, radial basis function networks and such

All such traditional neural networks assume that each input vector belongs to a single symbolic class, not an image which contains multiple objects in a cluttered scene. In this sense, all such traditional neural networks are symbolic, assuming that each input vector belongs to a single symbolic class. ART is symbolic in this sense.

> PDP connectionism actually starts with a predefined network, much like the models of AI.

I agree. Some, not all, PDP networks use handcrafted symbolic representations. Some nodes in the net are pixel like, but other nodes are symbolic classes. Those nodes that represent symbolic classes are not emergent. They belong to symbolic models that I defined.

Other PDP networks are emergent. Elman nets and Jordan nets are emergent, using error back-propagation. Their output does not have to be a symbolic class. The major problem of these two types of network is a lack of mechanism for transfer, like what finite automata can do but using handcrafted symbols. In addition, error back-propagation is not biologically plausible, since a baby does not have error at his effector ends.

> "At any rate, no model can get something from nothing and connectionist models are no exception. If this is what is meant by "emergent", then connectionism is not emergent, but neither is the brain or anything else."

The genome (Developmental Program, DP), is not emergent. The DP is not nothing. It is inherited. The brain is emergent, so is the entire body, except the first cell (zygote), which is inherited from the two parents. How do the body and the brain emerge? Autonomous development, regulated by the DP.

-John



On 7/22/11 12:12 AM, Juyang Weng wrote:
Christof:

I agree that the following unsupervised techniques are emergent:
SOM, local-to-global max in Creceptron, ICA (Independent Component Analysis), EM, Olshausen & Field's sparse coding criterion, and Poggio's HMAX (Poggio knew the local-to-global max in Creceptron and he seemed to like it).

However, such type of unsupervised learning seems almost never what any brain area is doing, not even LGN. Computationally, all supervised techniques are not scalable. The motor information must be used conjunctively with the sensory information. The brain anatomy says this loud and clear. The tough question is how. We did this way from 1994, through SHOSLIF, MILN, WWN, and DN, all of which are fully emergent.

-John



On 7/22/11 8:45 AM, Walter J Freeman wrote:
John,

Please send me a copy of your IJCNN 2010 paper about the 5-chunk brain-mind so I can have a look at it.

Walter



On 7/22/11 1:42 PM, Juyang Weng wrote:
Walter,

Here it is:
J. Weng, "A 5-Chunk Developmental Brain-Mind Network Model for Multiple Events in Complex Backgrounds,'' International Joint Conference on Neural Networks, July 18-23, Barcelona, Spain, pp. 1-8, 2010. PDF file.

I have a long version currently being reviewed by a journal. I quote from it:

"From the brain-mind model proposed here, it appears that a full understanding and appreciation of the (5+1)-chunk model for brain-mind requires depth and breadth of knowledge in at least six (6) disciplines, biology, neural science, psychology, computer science, electrical engineering, and mathematics. For example, computer science is not just a tool, but also a model for the biological brain. Steven Harnad 1990 [28] argued that the brain network must do arbitrary re-combinations like a computer. Neural network researchers, typically having an electrical engineering or physics background, did not figure out how a network can at least re-combine on the fly. The arbitrary re-combination is explained by immediate transfer in Theorem 3 below, in the sense that a new skill immediately re-combines with many new settings without explicit learning. The Finite Automaton (FA) theory is a brain model, but we must do it within a brain-like network. Interesting? Of course, more disciplines are related, such as philosophy, sociology, linguistics, laws, and politics. But the above 6 disciplines are minimally required. This requires new infrastructure for 6-disciplinary education beyond one's home discipline. "

The above shorter version saves your time.

-John



On 7/22/11 5:44 PM, jeff bowers wrote:
I think it is important to distinguish between “localist” or “grandmother” representations on the one hand, and “symbolic” representations on the other. The PDP approach in psychology denies both types of representations, and it has led to two parallel and ongoing debates in the cognitive sciences (that are quite distinct). Below is a passage from some unpublished work that highlights the difference.
Regarding ARTMAP – John is quite right that ARTMAP is supervised (as sometimes learning is supervised), but ART is not.

Passage regarding local vs. symbolic representations.
...The second question concerns the extent to which word knowledge is coded in a symbolic or non-symbolic format. According to symbolic theories of cognition, the long-term memory (LTM) representation of an item (e.g., a word, face, etc.) is coded in the same way in different contexts, whereas according to non-symbolic theories, the LTM representation of an item changes in different contexts. For example, a symbolic representation of the concept BROWN is invariant in different contexts, such as when thinking about a BROWN COW and a BROWN TREE (Fodor and Pylyshyn, 1988), whereas a non-symbolic representation of BROWN varies in these different contexts (e.g., Smolensky, 1988). The debate regarding symbolic vs. non-symbolic theories was initially highlighted in domain of high-level cognition, such as semantics (e.g., Fodor & Pylyshyn, 1988; Rumelhart & McClelland, 1986) and reasoning (e.g., Hummel & Holyoak, 1997; St. John & McClelland, 1990), but the distinction (and debate) applies to all research domains, including visual word identification and naming (e.g., Bowers & Davis, 2009; Sibley et al., 2009), morphology (McClelland & Patterson, 2002; Pinker & Ullman, 2002), and short-term memory (Botvinick & Plaut, 2009; Bowers, Damian, & Davis, 2009). For example, according to symbolic models of visual word identification, words are identified via letters that maintain their identity in different contexts (e.g., the D in DOG and GOD is coded by the same unit). Of course, letter order must be coded (in order to distinguish DOG from GOD), and this is achieved through a dynamic (temporary) process of assigning letters a position (cf., Bowers, 2002). By contrast, according to non-symbolic theories, words are identified via context dependent letter codes. For example, the LTM representation of D in DOG and GOD would be distinct (e.g., separate units for D-in-position-1 and D-in-position-3).

Unfortunately, the localist/distributed and symbolic/non-symbolic debates are sometimes conflated (with localist and symbolic theories treated as equivalent). However, these issues are orthogonal, and models can include all combinations of these representation types. For example, some models of visual word identification include local word representations and context dependent (non-symbolic) letter representations (e.g., McClelland & Rumhelhart, 1981), other models include local word representations and context independent (symbolic) letter codes (e.g., Davis, 2010), and still other (PDP) models include both distributed and context dependent letter and word representations (PDP models reject both local and symbolic representations1). The distinction between local and symbolic representations was highlighted by Hummel (2000) and Bowers (2002).
Jeff



On 7/23/11 5:04 AM, John Collins wrote:
Peter Foldiak wrote:

Horace Barlow's "Cell assemblies versus single cells" seems to be relevant here (particularly on the 'imaginary virtues of cell assemblies')

http://www.trin.cam.ac.uk/horacebarlow/157.pdf

Another paper that is important is Meyer & Damasio "Convergence and divergence in a neural architecture for recognition and memory",
Trends in Neurosciences Volume 32, Issue 7, July 2009, Pages 376-382 http://www.sciencedirect.com/science/article/pii/S0166223609000903

They make a distinction between a feature and a concept representation of the kind I mentioned a few days back. They have a concept called a convergence-divergence zone (CDZ), of which a grandmother cell is one possible implementation. A CDZ is "shaped by experience", i.e., is emergent. They make a distinction between a grandmother neuron defined as "a cell whose sole activity permits explicit mental images of objects and events", and one defined as "a cell whose activity correlates with the presence of a specific object". Only the second definition corresponds to a possible biological reality, and they intend that such a cell and its meaning are emergent. People like Jeff who argue for the possible existence of grandmother cells use the second definition. Those who argue for their impossibility use the first definition.

John



On 7/23/11 11:41 AM, Asim Roy wrote:
Hi, All,

The more I thought about the single cell recording experiments by Christof and his group, it became obvious that concept cells are at the cognitive level. These cells generate spikes that are interpretable and have meaning. So wrote up a few brain theories based on their findings. I think their findings are a major breakthrough. We finally may know how thought and cognition works, can actually track it down.

Anyway, take a look at the attached paper and let me know what you think. I would very much like to have arguments that can counter my ones. I am sure there are many.

Please treat this paper as confidential and do not distribute it to anyone else.

Asim



On 7/23/11 11:47 AM, Asim Roy wrote:
Here’s the paper. It’s easy to read. No mathematics. And it’s short. Would love to get your feedback.



On 7/23/11 9:27 PM, leonid wrote:
Dear Christoff, Asim, and everybody else:

any suggestion on how "concept cells" can contribute to automatically learning abstract concepts, such as "rationality" ?

Best,
Leonid



On 7/23/11 10:00 PM, Walter J Freeman wrote:
Dear John,

Thank you for your summary article. I attach mine herewith, from a lecture at IJCNN2009.

Our backgrounds are complementary with only partial overlap, so I can understand the gist of your model but cannot incorporate it into my own. Your model describes what brains do at higher levels of cognition than my animals can achieve. My models describe how brains do what they do, based on my observations of nerve cells at work in the execution of daily tasks that computers can't do. We have much to learn from each other.

Walter



On 7/24/11 1:20 AM, Juyang Weng wrote:
Asim,

I apply my prior arguments to your Fig. 1(a) as an example, as it might be clearer in your mind.

Fig. 1(a) in your writing supposes that the brain produces a symbol "cat". This is also a point I raised to ART: symbolic.

However, the brain never produces any symbol like "cat". The word "cat" is consensual in the mind of human communicators.

The "cat" may be the sound waves produced from your vocal tract, which requires many muscles to work together to make "cat".
Each time, your muscles work slightly differently when you pronounce "cat". You can verify that by looking at the sound waves or the firing patterns in your primary motor area. 95% human communicators may think that you said "cat" if they reach a consensus. But 5% of human communicators (e.g., babies and those over 80 years old) probably do not agree.

The "cat" can also be what you write on a piece of paper, but your hand writes slightly differently each time. The same principle applies.

Each muscles in your vocal tract also fire for other sounds, such as "car" and "dog".

This means there is no single "cat" cell in the sense that you drew in Fig. 1(a).

-John



On 7/24/11 1:55 AM, Asher Evans wrote:
Leonid,

It's interesting you should raise this question. I will be presenting a poster at IJCNN on Tuesday, Aug 2, conveying some of mine and John Collins's thoughts concerning this very matter among other things. I hope you and anyone else interested will stop by. The poster is #701 and is titled, "Modeling Knowledge Representation in Neuronal Networks."

I'm including our abstract below, but some of the key hypotheses are that:
* the brain ascribes "thingness" to stimuli by way of synchronous firing among neurons (or populations of neurons) representing the features and relationships that comprise the thing, cf. Wolf Singer, 1999 (Neuron)
* new concept neurons / populations emerge by linking to the set of synchronized items
* once these links are in place, the concept populations can then be linked to by still newer concepts
* population bindings have a short lifetime at first and become more established with recurrence

Cheers,
Asher (Garrett) Evans



On 7/24/11 2:31 AM, Asim Roy wrote:
John,

The concept “cat” has some physical form within the brain. You are arguing that it’s not in a single cell, but in a distributed form. Neuroscience might be the final arbitrar of this dispute. Note the discovery of triple invariant neurons, ones that are invariant to different images of an object and their written and spoken names. And it’s a single cell that is triple invariant!

In this paper, I am not exactly arguing for symbolic representation. The main argument is that there are high-level abstractions in the brain, whatever those abstractions are, encoded in single cells and that the spikes from these cells are interpretable and thus have meaning. Which means, these so-called concept cells are at the cognitive level. That definitely goes against the conventional wisdom that spiking neurons are at the subcognitive level and that spikes of individual neurons have no interpretation or meaning at the cognitive level.

So the real issue here is: Can spikes of individual neurons have meaning? Based on these single cell recordings, especially the experiments by Moran Cerf with “thought” tracking using a single neuron, my claim is that these concept cells have meaning on a stand-alone basis.

Asim




On 7/24/11 3:07 AM, Asim Roy wrote:
Hi Leonid,

Rationality is certainly a learned concept. Since it might take different forms in different situations, it might be context dependent (e.g. stock market investing vs. car racing or hunting). Whether it exists in a single concept cell, that’s a good question for Christof.

Asim



On 7/24/11 3:16 AM, leonid wrote:
Dear Asher, Great!

I'll visit your poster. You could look at my publications on the topic, including papers at IJCNN.

One difficult question, is how "abstract concepts" are identified for learning among zillions of combinations of simpler concepts.

Best
Leonid



On 7/24/11 3:20 AM, leonid wrote:
Agree! You exactly identified the difficulty.

Christoff, what you say?

Best,
Leonid



On 7/24/11 3:36 AM, Christof Koch wrote:
Hi Jeffrey,

sorry for not responding sooner.

All I can say that what the human MTL appears to contain is a very sparse (e.g. the vast majority of cells don't fire in response to the vast majority of images) and highly selective representation for concepts such as Jennifer Aniston or "E=mc^2" that the subject is very familiar with. How such cells develop can be explained using unsupervised learning algorithms that cluster the high-dimensional incoming data into a few & sparse clusters.

Using bayesian reasoning, we infer that maybe up to 10^6 neurons might be active in response to a very familiar concept and that each neuron might encode up to 50-100 such concepts. These are upper bounds. I suspect that the real numbers are much lower due to sampling bias of what we show, the problem of silent neurons, electrode bias etc. Based on biophysics, I suspect the number of cells that fire selective to any one concept might be more in the 1000 or less. But it's difficult to know without whole region coverage which can't be done in people but can be done in mice (thus, my recent move to the Allen Institute).

Yes, it is plausible that neurons fire progressively less the less the actual image the patient looks at relates to the concept 'represented' by that neuron.

Hope that helps

Christof



On 7/24/11 3:38 AM, Christof Koch wrote:
As I don't know what John means when he writes symbolic representation nor when Asim talks about cognitive representation, I refrain from using such labels. I know what we see at the end of our electrodes and that is remarkable sparseness and remarkable selectivity down to the level of individuals such as J. Anniston.

C.
----
Dr. Christof Koch



On 7/24/11 3:43 AM, Christof Koch wrote:
Are you asking for the concept of 'rationality'? That could be learned like any other frequently encountered pattern, via standard EM-type of unsupervised learning algorithm using something like a spare constraint. The idea is if you constantly think about 'rationality' (or J. Anniston or the Statue of Liberty or whatever else), your brain will tend to represent this complex and ever changing input pattern in a more selective & sparse representation. This probably happens in the MTL and is part of the necessary steps you need to go from attending to a particular pattern to consolidation of that pattern in long-term memory. That's what the electrodes in Dr. Fried's patients
are tapping into.

Christof



On 7/24/11 4:34 AM, leonid wrote:
The difficulty is what rationality is? and why one would think about rationality or gibberish (jibberish... or any other meaningless selection of simpler ideas). It gets easier after one tries to understand something, but what is worse understanding? Consider a "simpler case" of learning situations, which could be modelled, for simplicity, as collections of few "important" objects. The fundamental difficulty is that these important objects are immersed in zillions of unimportant objects (say a scatch on the wall is unimportant for understanding "professor office"), and "situations" are immersed in zillions of meaningless combinations of objects. (Situation learning has remained mathematically unsolved after 50 years of attempts.)

Would you say that your findings are consistent with the following (much simplified) view:

Bottom-up projections from retina onto visual cortex are matched there with top-down projections from "memory". After matching, an object is recognized by your "concept cell" at MTL. This is a simplified description of Bar et al 2006 perception experiment - except for your concept cell.

Or would you suggest that your findings are different in principle and represent an alternative perception mechanism?

Best
Leonid



On 7/24/11 4:41 AM, Christof Koch wrote:
Well, Leonid, that simple explanation of your's is compatible with a lot of different findings, including ours. The question I'm interested in is how such concept cells develop
C.
----
Dr. Christof Koch



On 7/24/11 6:44 AM, leonid wrote:
This is a good question. I am doing mathematical modeling, not experimental evolutionary neuroscience. Math model can help sometimes.

I am glad that your work is compatible with what I know. It seemed some participants in this discussion erroneously thought that your "concept cell" is the entire mechanism of thinking.

Best
Leonid



On 7/24/11 7:08 AM, Christof Koch wrote:
For heaven's sake no. it's just one component of an immensely complex system

;-)
----
Dr. Christof Koch



On 7/24/11 9:01 AM, Juyang Weng wrote:
> the concept BROWN is invariant in different contexts, such as when thinking about a BROWN COW and a BROWN TREE (Fodor and Pylyshyn, 1988)

(Fodor and Pylyshyn, 1988) were incorrect. Like many psychologists, they did not look into the brain carefully, although by 1988 many studies in neural anatomy have already published.

I repeat: There is no neuron in the brain that has the iff (if and only if) condition about the BROWN concept.
In fact, BROWN means different things in different brains, and even at different times and different contexts. I hope that my prior example about CAT applies to any extra-body concept, including BROWN. Our WWNs schematically explained how the brain learns concepts in cluttered environments. The same scheme should apply to any human communicable concepts. WWNs do not need any concept cell or any grandmother cell inside its internal area to do almost perfectly as long as they have enough number of neurons.

-John



On 7/24/11 9:19 AM, Juyang Weng wrote:
Dear Walter:

You wrote: "In some beats that recur at theta rates (3-7 Hz), the order parameter transiently approaches zero, giving a null spike in which the microscopic
activity is uniformly disordered (symmetric)."

This is consistent to our model. The brain cannot sort out the winners quickly. It can only iterate through back-and-forth excitations and inhibitions.
Each neuron cannot generate two consecutive spikes too quickly. The peak spiking rate around 120 spikes/second has been generated only in lab conditions.

-John



On 7/24/11 9:27 AM, Juyang Weng wrote:
Asher,

I do not see major problems with your vague "key hypotheses". Computational modelers face the problem of making all these vague ideas explicit and show them in an all round theory (proofs) and computer simulations, including spatial attention, temporal attention, cluttered background, incremental learning, neuromorphic computation (Christof: the EM way does not seem to fit neuromorphic computation) and so on. This is what I wrote: "A tough question is how". You did not mention how the brain deals with cluttered backgrounds.

-John



On 7/24/11 9:36 AM, Juyang Weng wrote:

>On 7/23/11 5:04 AM, John Collins wrote:
> People like Jeff who argue for the possible existence of grandmother cells use the second definition.
> Those who argue for their impossibility use the first definition.

John, good. Then, I personally think "correlate" is too weak for the term "concept cell", since one neuron "correlates" with many many other concepts. Since this is a free-of-speech country, let us define what is meant when we use a term like "concept cell". Not just "a cell correlates with one concept", but also "a cell correlates with many other concepts".

-John


On 7/24/11 9:42 AM, Juyang Weng wrote:
Weng IJCNN 2010 explained
(1) bottom-up input is only part of the causality of a brain cell. Top-down input from motor and lateral input from other cells in the same and other related area played a major role.
(2) If the above is true, any human communicable concepts such as "rationality" can be learned by the WWN model, from a cluttered background. A generative version of WWN does not need the iff condition for its internal cells in order to learn immediately, error free, for training examples.

-John


On 7/24/11 9:54 AM, Juyang Weng wrote:
Christof:

You wrote: "I suspect the number of cells that fire selective to any one concept might be more in the 1000 or less". I assume that you meant "in MTL".
This is consistent with our DN brain-mind model.

You wrote: "How such cells develop can be explained using unsupervised learning algorithms that cluster the high-dimensional incoming data into a few & sparse clusters."
This is inconsistent to the brain anatomy and your DN model. Every cell has a top-down part. Almost no cell is totally unsupervised (exception: retina in primate, but not other mammals). Every cell is supervised from motor, directly or indirectly. The supervision can be self-practice. When WWN practices it learns too. That is, self taught.

-John



On 7/24/11 10:02 AM, Juyang Weng wrote:

> On 7/24/11 4:34 AM, leonid wrote:
> The difficulty is what rationality is?

Leonid, rationality is different in every different brain. The DP of a brain does not "care" about this question, as every brain is different. A concept is "consensual" among a human group. For example, "rationality" in your brain is slightly different from "rationality" in my brain. We should consider "rationality" like any other human communicable concepts when we understand how the brain-mind works. For example, "brown" is slightly different in different brains.

-John


On 7/24/11 10:16 AM, Juyang Weng wrote:

> On 7/24/11 3:16 AM, leonid wrote:
> One difficult question, is how "abstract concepts" are identified for learning among zillions of combinations of
> simpler concepts.

In our DN model, a more "abstract concept" and a "simpler concept" are treated basically the same by the DP.
Young children have problems in learning more "abstract concept" (e.g., numbers) before they have learned many "simpler concepts" (e.g., car, apple).
For example, the learning of abstract concept like "number 2" needs support of skills to classify objects (e.g., car, apple) from a cluttered background.

Kelly Mix at MSU and Linda Smith at Indiana University have plenty of such examples. Our MILN has shown that how learning of simpler concept "numbers" first helped it to learn more abstract concepts such as "odd number" and "even number". See Fig. 7 in the following paper:

J. Weng, H. Lu, T. Luwang and X. Xue, ``Multilayer In-place Learning Networks for Modeling Functional Layers in the Laminar Cortex'' Neural Networks, vol. 21, no.2-3, pp. 150-159, 2008. PDF file.

-John


On 7/24/11 10:40 AM, Juyang Weng wrote:

> On 7/24/11 6:44 AM, leonid wrote:
> some participants in this discussion erroneously thought that your "concept cell" is the entire mechanism of
> thinking.
> Leonid,

I raised that the bottom-up explanation of Christof is not sufficient for the development of even "feature cells", such as cells in V1, let alone cells in MTL.

However, in IJCNN 2010, I have explained how DN general model of an internal cell inside the brain explains some major mechanisms of "the entire mechanisms of thinking". Note, "the entire mechanisms of thinking", not singular.

In IJCNN 2011, I will explain it further: immediate and error free, along with a sketch of mathematical proof.

I suggest that neural network researchers learn automata theory in computer science, as such knowledge background is useful for understanding how the brain thinks. For example, the automata theory background is necessary to understand how the brain does transfers: Transfer all un-modeled concepts learned in one setting to many other settings without a need for explicit learning. I have not yet seen any other neural networks that have reported such a capability.

When I wrote brain-mind model, I did not claim all brain-mind functions. However, I do not write brain-mind lightly. For example, there seemed to be no prior neural network model that has explained how a spatial network can do general-purpose temporal processing. Temporal processing is just one chunk among the 5 chunks of our DN brain-mind model.

-John



On 7/24/11 1:39 PM, Juyang Weng wrote:

> On 7/24/11 10:40 AM, Juyang Weng wrote:
> I raised that the bottom-up explanation of Christof is not sufficient for the development of even "feature cells",
> such as cells in V1, let alone cells in MTL.

To put in another way, top-down signals directly or indirectly from the motor area are critical for the emergence of the highly selective cells in MTL that
Christof and others reported (but they are not "iff" concept cells).

-John



On 7/24/11 1:55 PM, Christof Koch wrote:
Well, that's an interesting suggestion that could, at least in principle, be tested using
optogenetics.

C.
----
Dr. Christof Koch
On 7/24/11 2:02 PM, Juyang Weng wrote:
Yes, optogenetics studies about this neuroanatomy-based theoretical prediction could be a step further from neuroanatomy.
Christof, let me know if I can be of any assistance.

-John



On 7/24/11 3:20 PM, Asim Roy wrote:
Hi Christof,

Would it be fair to say that a concept cell that we might think of as “representing” Jennifer Aniston is just a very high level abstraction that is closely related to Jennifer Aniston, but also to other lookalikes. I guess, from your experiments so far, that’s all one can really claim, that there is a high level abstraction in these concept cells that are tightly related to a small set of objects. However, with further experiments, once you are able to monitor whole regions of the brain, you might be able to make stronger claims about “representation” by concept cells. Let me know if that’s a fair assessment.

Asim



On 7/24/11 3:41 PM, Asim Roy wrote:
Hi Christof,

In the Moran Cerf experiments, from the enhanced spikes of a particular neuron, you were able to determine which of the two images the patient was thinking about. The patient didn’t have to say “I am now thinking about the target image” or “I am now thinking about the distractor image.” You knew which concept (image) the patient was thinking about from the spikes of the associated neuron. That makes the spikes of those particular neurons interpretable and have meaning to us. That’s all I meant by saying that these spikes of single neurons are at the cognitive level and have meaning. I think that was a great experiment.

Asim



On 7/24/11 4:18 PM, Asim Roy wrote:
“Cognitive level” only means the spikes are interpretable, as in the case of the Cerf experiments. Why doesn’t it “mean much” in these experiments? The experiments were great. And why would one need iff on interpretations? Interpretations can vary, but they are still interpretations and have meaning.



On 7/24/11 8:48 PM, leonid wrote:
John,

I suggest that it is imperative to analyze the process of natural learning (without training) from the point of view of its computational complexity. In the 1950s, many scientists thoughts that computers will soon surpass humans in creativity. After 60 years we know that most algorithms, which seemingly could learn everything, failed because of computational complexity.

Best,
Leonid




On 7/24/11 8:52 PM, leonid wrote:
John,

Of course you are right. But how this can be modelled mathematically, while avoiding exponential explosion of learning?

Best
Leonid



On 7/24/11 8:58 PM, leonid wrote:
John,

I agree with the second part of your statement about objects. But I emphasize the difficulty of how a child (without training) would know that it is important to understand an abstract concept ("number2", or "dining room") and that it is not important to understand a random collection of objects (like a pattern on the floor, a scratch on the wall, a piece of garbage) as an "abstract concept" important for learning and understanding.

Best
Leonid



On 7/24/11 9:05 PM, leonid wrote:
John,

I am looking forward to getting together and exchange ideas. When I criticized excitment about "concept cell" explaining everything about the mind, I did not mean specifically you at all, I referenced the collective tendency, despite Principe, Freeman, Coch, and some others cautioned against staking too much on a single mechanism.

Best,
Leonid



On 7/25/11 2:54 AM, bruno apolloni wrote:
Dear Asim

I did read your stimulating paper. On some parts I agree, on other parts I disagree. But it is a so big matter that I doubt to be able to synthesize my feedback in a short mail. Rather, why not using the paper as a trigger for an interesting discussion during the AML SIG meeting? In this case I could prepare a couple of slides.
Considering the huge exchange of mails on the grandmother cell, I assume that other colleagues would like to do the same. Thus, we could have a panel on this matter.

What you think about?

Best

Bruno



On 7/25/11 2:50 PM, Asim Roy wrote:
Here’s a summary of the Moran Cerf experiments (Cerf et al. 2010) from my paper:
“In their experiments, twelve epilepsy patients played a game where they controlled the display of two superimposed images. The controlling was done through the activity of four MTL neurons. Before the actual experiment, the researchers identified four different neurons that responded selectively to four different images. In these experiments, one of the four images was randomly designated as the target image. Each trial started with a short display of a random target image followed by an overlaid hybrid image consisting of the target and one of the other three images (the distractor image). The subject was then told to enhance the target image by focusing his/her thoughts on it; as per Cerf et al. (2010), the patients were instructed to ‘‘continuously think of the concept represented by that image.” The initial visibility of both the images was 50% and the visibility of an image was increased or decreased every 100 ms based on the firing rates of the four MTL neurons that were monitored. Spike counts were actually used to measure firing rates and to control the visibility of the two images on the screen. In general, if the firing rate of one neuron was higher compared to the other, more visible was the image associated with that neuron and less visible the other image. The trial was terminated when either one of the two images, the target or the distractor image, was fully visible or after a fixed time limit of 10 seconds. The subjects successfully reached the target in 596 out of 864 trials (69.0%; 202 failures and 66 timeouts).”

Where is the “imprecision” in the implied interpretations by the researchers in these experiments? They have directly linked the enhanced firings of a particular neuron with “thought” about the corresponding image? They never had to ask the patients: “What image are you thinking about?” They knew it from the firings of the four monitored neurons. It doesn’t really matter if these neurons respond to other closely related images. It doesn’t have to be a unique response (if and only if type that you are fixated on). A concept cell could respond to Jennifer Aniston or some lookalike. It will still have “an interpretation and a meaning.”

And it doesn’t matter how it is expressed in human language, whether one says “I am thinking about image A” or “My brain is now focused on image A” or some other variation of it. It still has “an interpretation and a meaning.” And Moran Cerf did the experiments with that understanding.
Asim



On 7/25/11 9:00 PM, leonid wrote:
Asim,

You are right, the problem is tough. Problems of modelling intelligence have not been solved mathematically for 50-60 years, except in "toy" examples, because of this problem. Several times I published papers analyzing large classes of popular algorithms and demonstrating that complexity is exponential on average. I have some success with developing linear complexity algorithms for several important real-world problems, inspired by what we know about the mind, and even predicting results of future mind experiments (some confirmed), but there is no general mathematical results for conditions of linear convergence.

You are right, one of the problems is that most researchers are not aware of the complexity problem. Every year hundreds to thousands of various algorithms are published with demonstration on "toy" problems, the algorithms "look like" they should be able to learn any solution to any problem that the mind can solve; and their authors think that the problem is solved. But as soon as real-world learning without a teacher is approached, most algorithms fail. E.g., explaining high level cognition and its interactions with language is not even attempted (and rightly so) - I would suggest that we all should be at least aware of the complexity problem. There are fundamental difficulties similar to Godelian difficulties with logic. It would be good if more researchers would be aware of this and of the need for fundamental arguments of how to overcome the complexity.

Learning without a teacher is fundamentally combinatorially complex: because interactions betwen top-down and bottom-up signals are combinatorially many.

There are general ideas of the prolems and solutions, but too few researchers are aware. This was my reason to include this as a topic.

Best,
Leonid



On 7/25/11 10:32 PM, Juyang Weng wrote:
> We do not have mathematical models for any of this

Yes, we do. The DN model. But there is no need to model any extra-body concepts, let alone "the entire culture".

> How language interacts with cognition - a fundamental problem - are there any papers on this topic?

Yes, DN did it and published:

K. Miyan and J. Weng, ``WWN-Text: Cortex-Like Language Acquisition with ’What’ and ’Where’,'' in Proc. IEEE 9th International Conference on Development and Learning,'' Ann Arbor, pp. 280-285, August 18-21, 2010 . PDF file. (text as perception, early language learning and early language generalization.)

J. Weng, Q. Zhang, M. Chi, and X. Xue, "Complex Text Processing by the Temporal Context Machines," in Proc. IEEE 8th International Conference on Development and Learning," Shanghai, China, pp. +1-8, June 4-7, 2009. PDF file. (Temporal text.)

In other words, language as one of many emergent capabilities, embedded as perception, cognition and behaviors within a tightly integrated brain-like network.

-John



On 7/25/11 10:41 PM, Juyang Weng wrote:

> On 7/25/11 9:00 PM, leonid wrote:
> there is no general mathematical results for conditions of linear convergence.

Leonid, I was having the same concerns as yours about 5 years ago. The situation has changed. DN is immediate (there is no iterative convergence problem) and has a linear time complexity in the number of neurons in the network. For real-time online system, the complexity issue has an interesting brain-like "new world".
The brain does face the same complexity issue of symbolic algorithms.

-John



On 7/25/11 10:58 PM, Juyang Weng wrote:
Asim,

I thought about that interesting experiment when I read that paper. At that time, at least I could schematically explain how a DN can perform this experiment and be successful if some simplifications are made (e.g., simplify the natural language communication part). However, the meanings of each cell in the DN cannot be explained precisely in any human language.

-John



On 7/25/11 11:04 PM, Juyang Weng wrote:
Agreed, Leonid. Many mechanisms must work together very tightly and, in fact, beautifully. -John



On 7/25/11 11:11 PM, Juyang Weng wrote:
Your emphasis needs many mechanisms to accomplish, almost all in the brain. However, one important factor is what the child does, which "tells" the brain which parts in the environments are "important". The brain needs synapse maintenance to accomplish that, whose model will be presented by my student at IJCNN 2011.

-John



On 7/25/11 11:23 PM, Juyang Weng wrote:
> But how this can be modelled mathematically, while avoiding exponential explosion of learning?

"Suppose that an agent, natural and artificial, needs to deal
with c concepts and each concept takes one of v values.
The corresponding symbolic model potentially requires v^c
different states, exponential in c. Letting c = 22 and v = 4,
the symbolic model potentially needs v^c = 4^{22} = (4^2)^{11} =
16^{11} > 10^{11} = 100; 000; 000; 000, larger than the number of
neurons in the human brain. Here are 23 examples of concept:
object type, horizontal direction, vertical direction, object
pose, apparent scale on the retina, viewing distance, viewing
angle, surface texture, surface color, surface reflectance, lighting
direction, lighting color, lighting uniformity, material,
weight, temperature, deformability, purpose, usage, owner,
price, horizontal relationship between two attended objects,
vertical relationship between two attended objects. This is a
complexity reason why the brain cannot use symbolic states
for its internal representations."

"Next, consider the WWN. For c concepts, each having v
values, the number of motor neurons in WWN is only vc.
With v = 4 and c = 22, vc = 88 only, instead of 16^{11}."

The above is extracted from Weng, "Symbolic Models and Emergent Models: A Review" to appear in TAMD. I am afraid that the above "quote" is not very clear, but it gives you an idea about how. The main power lies in the distributed representation (in motor area in this case).

-John



On 7/26/11 4:20 AM, Asher Evans wrote:
Dear John,

> Letting c = 22 and v = 4, the symbolic model potentially needs v^c = 4^{22} = (4^2)^{11} =16^{11} > 10^{11} = 100; 000; 000; 000, larger than the number of neurons in the human brain.

I agree with you that this is a highly relevant concern. To make sure we're on the same page, allow me to use an example. Let's say we are talking about a house. Certainly there are at least 22 parameters for the house, including square footage, height, type of floor, age, number of bathrooms, number of appliances, location, etc. with at least four values for each. If I am understanding you correctly, you are saying that in order to represent such a house using concept neurons, it is necessary to have a separate concept population for each combination of values of the 22 parameters, which is indeed 4^22 (v^c), far too many.

Binding by synchrony can drastically reduce this. Suppose that one has a concept population for each of the values of the c parameters, independently (v*c populations). The house can be represented by synchronous activity among a select set of c of the populations. The synchrony represents the fact that this set of parameters goes together as a single entity, this particular house, which has these particular parameter values simultaneously.

There are cases for which v*c is still undesirably too many particularly when v gets very high as it does for the different dimensions of spatial location and separation, but there is a way binding can reduce this as well by linking generalized number concepts, which humans clearly have, to concepts standing for the parameters. If you would be interested in stopping by my poster next Tuesday, I would be happy to explain in greater depth and detail.

I agree that simulation is a crucial stage in developing and verifying an idea or set of ideas. John C. and I are of the mind, though, that the thinking we have done so far is worth sharing. Apparently the abstract selection folks agree.

Best,
Asher



On 7/26/11 5:23 AM, Asim Roy wrote:
It’s a wrong analysis. You don’t need to hold those values at the same time. You just explore cleverly the promising solutions. Greedy heuristics are part of all major solution techniques when you have such astronomically many possibilities to explore. That’s what done by all optimization techniques.

Asim


On 7/26/11 5:52 AM, Asher Evans wrote:
Dear Leonid,

Forgive me for not recognizing or looking up the name. I'm actually signed up for your Dynamic Logic tutorial at this year's conference. I was looking at your paper from IJCNN 2008 just now, which I do plan to read in depth, and I'm looking forward to understanding your ideas better with the opportunity to learn about them in person.
Thanks for your comments; I agree this is a critical question, and we have thought about it. I'm very much looking forward to discussing in depth with you at IJCNN.

Cheers,
Asher



On 7/26/11 8:54 AM, Harry Erwin wrote:
So single-presentation learning should not be associated with concept cells. What do the experimental data say about that?

On 25 Jul 2011, at 15:17, Christof Koch wrote:
You are entirely correct Asim. A pattern needs to show up again and again in the relevant neural networks (here in the human MTL) for 'concept cell' to be formed

C.
----
Dr. Christof Koch

--
Harry Erwin



On 7/26/11 9:06 AM, Christof Koch wrote:
We find category cell to novel stimuli but never selective neurons. We do see the emergence of such cells after a few days (e.g. we have a few such invariant and abstract cells responding to the Caltech personnel that the patient interacts with for a few days).

C
----
Dr. Christof Koch



On 7/27/11 10:49 PM, leonid wrote:
Asim,

You are right about most algorithms. I know of one, dynamic logic, that is not greedy, not exponentially complex, often achieves global maximum, works practically, exceeds other algorithms for difficult problems by orders of magnitude, prooven as a good model for brain mechanisms of perception and cognition (unique in predicting vague-to-crisp process for representations - confirmed in experiments), and suggests explanations for several mysteries of language-cognition-music, which have not been modelled by other algoritms (e.g., why kids learn language first, and cognition later? how words are associated with objects? how explain real existence of free will, and what is it? what is a function and origin of music?). These hypotheses are being tested in several psychological and neuroimaging labs.

Best,
Leonid



On 7/27/11 11:02 PM, leonid wrote:
Dear Asher,

It is even much more difficult, to learn what is "house" among zillions of surrounding objects. Of course everyone makes "weighted"evaluation on all features. But how do you autonomously learn what features even to consider? Your example is highly constrained by prior knowledge, where does it automously come from?

Looking forward to meeting at IJCNN,
Best,
Leonid



On 7/28/11 3:17 AM, Asim Roy wrote:
Leonid,

That sounds great. Please send me a copy of the paper.

What I meant was that no technique holds in memory all the possible solutions. You iteratively generate and explore promising solutions and find the optimal or, in the case of heuristics, find whatever “best” solution you can find. That’s all I meant. I am sure your dynamic logic algorithm works the same way.

Asim




On 7/28/11 7:07 AM, Christof Koch wrote:
There i no need to exhaustively show all possible permutations along all possible directions. Unsupervised learning in machine vision learns from the real world where
many constraints exists such that 'nature' scenes (e.g. houses or faces) have many, independent contraints. So you can learn from showing the system a couple of 100s
or 1000s houses or faces or chairs...

C,
----
Dr. Christof Koch



On 7/29/11 8:18 AM, Asim Roy wrote:
We have been debating single cell recordings and concept cells for over a month now. The more recent experiments on tracking “thought” by monitoring spikes of single neurons were done by Dr. Moran Cerf at UCLA medical school with epilepsy patients of Dr. Izthak Fried. We are lucky that he could take time off from his experiments to give a talk at our Autonomous Machine Learning (AML) SIG meeting on Tuesday, August 2, evening. We will start the meeting at 8 pm in San Juan so that you have a chance to meet and chat with him. His actual talk will start at around 8:45 pm to minimize overlap with the poster sessions. Dr. Cerf’s short bio and abstract for the talk are below. I want to thank Ali Minai for facilitating this visit. And, of course, Christof Koch for arranging the talk since he will be in Japan next week.
Please publicize this talk. I believe this group has some breakthrough findings about the brain that has the potential to change our thinking about how the brain works and learns. Dr. Cerf will also talk about his more recent experiments which have not been published yet, including “how memory and perception interact in the human brain; and show some of the recent results from unpublished works related to emotions and free will in humans.”
There will be plenty of time for Q&A. Hope to see you on Tuesday evening for this interesting talk.
Asim Roy



On 7/29/11 9:21 AM, leonid wrote:
Asim,

Here are few papers. There are several ways to understand dynamic logic, I prefer to understand it as a process from vague to crisp, with vagueness of models corresponding to uncertainty of parameters and associations (of bottom-up and top-down signals; in other words, associations between models and data). If you look at the figures in these papers, you see right away that it is not similar to other algorithms, it is fundamentally different. This fundamental difference is in that the process you described is limited by Godelian limitations, which dynamic logic overcomes.

The first two papers demonstrate 100 to 1000 times improved engineering solutions, and how they relate to cognition. The 3rd paper suggests a hypothesis of how cognition interacts with language. The last paper (WMC) has no new engineering solutions, but a cognitive model - a mathematical formalization of Barsalou's Perceptual Symbol System, a mathematical explanation of what is a symbol (not a static sign of AI, but a dynamic process connecting unconsciousness to consciousness), and how symbols in the brain become grounded.

The reason I send several papers and explain in details the meaning of dynamic logic is that I am looking for collaborators, for researchers interested in unexplained mysteries. For example, what is a function of the beautiful in cognition (it is not sex), what is a function of music in cognition - this was called a mystery by Aristotle, Kant, Darwin, Pinker (also Steve does not really know such a word). If you know other algorithms explaining function of music in cognition (which would be acknowledged by psychologists, cognitive musicologists, and mathematicians - please let me know). There are dozens of unsolved mysteries, hundreds of dissertations to be written, and research funds are available - so I am looking for collaborators interested in fundamental questions.

Asim, your search for fundamental issues creates opportunities for real discussions, and this is most precious in science. Best,
Leonid



On 7/29/11 2:09 PM, Juyang Weng wrote:
Asim,

You will lead to a brittle system as many in symbolic AI have done, regardless how you "explore cleverly the promising solutions" using symbolic representations.

We must put into math to be clear:

Let us look into the spatial problem only first. Consider an open number of objects $\{ o | o \in E\}$ that a system will run into, where $E$ is the external environment. Each object $o$ has $c=22$ symbolic concepts, each concept having $v=4$ values. For example, $o=(v_1, v_2, v_2, ..., v_{22})$, where $v_{i}
\in \{do-not-care, low, high, care-but-missing-information-in-the-data\}$, $i=1, 2, ... , 22$.

The required output action $z$ from the agent depends on the properties of the current object in an arbitrary, but in task-specific way. $z=1$ if $o$ is acceptable by the task-agent. $z=0$ if $o$ is not acceptable. That is, $z=f(o)$, were $f$ represents the classification task-agent. Using symbolic representation, there are $v^c$ nodes (i.e., your ``concept cells'') in the decision space, called version space in AI. There are $v_c=4^{22}$ symbolic nodes to classify by $f$.

Try to run ART on this problem, as ART takes a vector $o=(v_1, v_2, v_2, ..., v_{22})$ as input to learn. How many vectors $\{o\}$ does ART need? How many nodes does ART need to learn the task-agent correctly?

Of course, in symbolic AI, the designer tries to reduce the number of $4^{22}$ possibilities, by static handcrafted design: Drop some nodes, merge some nodes, to
form equivalent classes of nodes $A_i$, $i=1, 2, ... , n$. If the number of equivalent class $n$ is small enough, learn $n$ data points only. A major problem arises: In the real world, many nodes inside a hypothetically equivalent class $A_i$ require different $z$. The human designer cannot examine all $4^{22}$ nodes to determine the equivalent classes $A_i$, $i=1, 2, ... , n$, since there are more than 100, 000, 000, 000 of them!

This is my math explanation about why symbolic systems are brittle in the real world.

This is also the problem with any unsupervised learning, if symbolic nodes are used, Christof.

-John



On 7/29/11 2:16 PM, Asim Roy wrote:
Let’s talk about this at IJCNN. That’s what meetings are for.



On 7/29/11 2:12 PM, Juyang Weng wrote:
Good, Christof. The cells must emerge from interactions but also use motor neurons (e.g., saying the name of the object). That is how DN does it.

-John



On 7/29/11 3:32 PM, Juyang Weng wrote:
Leonid:

I read your papers before when you sent me papers. Your model gives a way to measure the match between input vector and the prediction (parameter vector) of each model. Let me know if I am correct in my current guess below.

Your NMF model has the following properties.
(1) Your similarity measure: $L({X}, {M}) = \prod_{n\inN} \sum_{m\in M} r(m)l(X(n)|m)$ is a match between sensory input $X$ and one or more models in $M$
based on a single parameter vector $S_m$ which is not explicit in the above expression.
(2) Your iterative procedure is similar to ART in that it is for finding hopefully the "best" matched model among multiple candidates. ART does this iterative search process one model at a time (one model produces one top-down expectation). Your model does this iterative process through a numeric way: along partial derivatives.
(3) ART makes sure that only one model parameter vector is active at a time using logic-like control; your model searches for a single set of model parameters (position, orientation, or size of an object, potentially 22 or more) through gradient-descent. ART allows one model to have one parameter vector only. Your
formulation in (1) potentially allows one parameter vector to be "credited" by multiple models.
(4) Your model is not neuromorphic: The system does not consist of a network of uniform neurons where every neuron has a synaptic weight vector and learns using biologically known mechanisms.

-John




On 7/29/11 3:51 PM, Juyang Weng wrote:
Dear Leonid,

Yes, you are correct. "It is even much more difficult, to learn what is "house" among zillions of surrounding objects." DN deals with such difficult problems. Furthermore, DN learning is fully autonomous inside its "brain".

-John



On 7/29/11 9:09 PM, Juyang Weng wrote:
Dear All,

To facilitate our continued discussion at IJCNN 2011, your email contributions are retrievable from our IJCNN Panel Session web page:
Panel Session
Brain-Mind Architectures: Module-Free, General Purpose, and Immediate Learning?
http://www.cse.msu.edu/ei/IJCNN11panel/

You can also find this link from my home page.

See you all at San Jose, Sunday July 31, 2011.

Best,

-John

Top | Abstract | Organizers | Tentative Panelists | Email Discussions

 Back To Weng's Home Page: http://www.cse.msu.edu/~weng/