Panel session for IJCNN 2010
There have been a growing interests in simulating brain functions at the brain scale. However, there is a large gap between connectionist modeling and symbolic modeling.
Some researchers said: “Neural networks cannot do reasoning.” This was negative, but it pointed out a wide knowledge gap about the brain. This situation has created a major credibility problem of all connectionist approaches to understanding the brain and to simulating the brain-like capabilities.
Connectionist approaches are bottom-up (e.g., from pixels) and symbolic approaches are top-down (e.g., from abstract concepts). Between the concrete (e.g., an edge or an edge grouping) and the abstract (e.g., goal), “much in-between” is missing. What is the “much in-between”?
This panel will bring active researchers who are interested in this subject to discuss and debate the related open questions.
Those who have substantial ideas on this subject and like to participate as panel members are encouraged to send an email to the two organizers.
This panel is co-sponsored by:
IEEE CIS Autonomous Mental Development TC and its Visual Processing Task Force
INNS Autonomous Learning SIG
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
1. Does the brain use symbolic representation in a fashion similar to our symbolic AI? Why? What lesson can we learn?
2. Connectionist models have shown progress in pattern recognition, but they have been largely used as a classifier or regressor. Some said that artificial neural networks cannot perform reasoning. Is there "a light in the tunnel" for connectionist models to perform goal-directed reasoning? Why?
3. How do you think about the brain/artificial architecture for the autonomous development? As the brain is "skull-closed", how does it fully autonomously develop its internal representation from one task to the next? What is the “much in-between”?
Prof. Paolo Arena, University of Catania, Italy, parena@diees.unict.it
Prof. Minoru Asada, Osaka University, Japan, asada@ams.eng.osaka-u.ac.jp
Prof. Angelo Cangelosi, University of Plymouth, UK, acangelosi@plymouth.ac.uk
Prof. Nik Kasabov, New Zealand, nkasabov@aut.ac.nz
Prof. Giorgio Metta, University of Genova, Italy, pasa@dist.unige.it
Prof. Asim Roy, Arizona State University, USA, asim.roy@asu.edu
Prof. Ron Sun, Rensselaer Polytechnic Institute, USA, rsun@rpi.edu
Dr. Narayan Srinivasa, HRL, USA, nsrinivasa@hrl.com
Prof. Janusz Starzyk, Ohio University, USA, starzyk@bobcat.ent.ohiou.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
Dr. Paul Werbos, National Science Foundation, USA, pwerbos@nsf.gov
-------- Original Message --------
Date: Fri, 02 Apr 2010 10:53:33 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: asada <asada@ams.eng.osaka-u.ac.jp>
CC: pwerbos@nsf.gov, acangelosi@plymouth.ac.uk, nkasabov@aut.ac.nz, pasa@dist.unige.it, ASIM.ROY@asu.edu, rsun@rpi.edu, taymore2002@aol.co.uk, parena@diees.unict.it, nsrinivasa@hrl.com, starzykj@bobcat.ent.ohiou.edu, mcclelland@stanford.edu
Subject: Kick-off email: Introduction
Dear All,
I would like to thank Ron Sun and Minoru to take the lead in submitting their first version of slides. A panel session will not have sufficient time to discuss and think about these questions. Note, everybody can update their panel slides as often as he likes. Let us do email discussion first.
To start discussion and keep past references I will post our emails on our web, as we did for NSF/DARPA funded Workshop on Development and Learning (WDL): http://www.cse.msu.edu/dl/email.htm
Please book mark our IJCNN 2010 panel web site: http://www.cse.msu.edu/ei/IJCNN10panel/
To be easier for others to read, I will simply append emails one after another, in the chronic order. Simply scroll down to get the latest emails.
If there is any objection or suggestion, please make a motion and we will discuss, move and vote among our panel members.
I will start the email discussion first. To be clear, our discussion should be direct. Please do not get upset. We only address the scientific and technical issues for this panel, not personal characters. I will serve as the moderator. If there is any personal attack, I will cut that part and leave [PA] as a place left out as personal attack before posting the remaining parts.
Best regards,
-John
-------- Original Message --------
Date: Thu, 01 Apr 2010 20:27:30 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: suppressed
Subject: A brain-mind picture: Your expert's vision, comments, and suggestions are needed
Dear respected domain experts:
In many ways, I was your student and I still am.
Narendra Ahuja and Tom Huang taught me computer vision. Lawrence Rabiner's work taught me how HMM deals with time warping. Tommy Poggio's work attracted me into computational neuroscience. Mriganka Sur taught me developmental neuroscience. Peter Shiller showed me much detailed visual brain. Michael Merzenich's work made me think hard about how the brain deals with time. I would like to say "sorry" to Robert Desimone for not matching his kind face with his name when Mriganka Sur introduced me to him during an ice cream break at MIT. Robert Desimone's papers taught me so much about the miraculous attention that the brain does. Marvin Minsky told me a cold fact: artificial neural networks cannot reason. Alan Leshner made very useful suggestions. Almost everybody here taught me something important. That is why I write to you.
Now I am scared. Nobody seems to be able to confidently evaluate the picture that I piece-together. The scientific history seems to repeat itself again! This is what the first short perspective article (file 1, name suppressed) is about.
I am "crazy" and "naive" enough to seek for the big truth, without regret for the sudden stop of the flow of all my federal funding since 2002. To save your time, a brain-mind picture is greatly simplified into a single journal page in the second attachment (file 2, name suppressed).
If you want to gain a more complete picture, please feel free to browse the third attachment (file 3, name suppressed).
All the 3 attachments are unpublished manuscripts under review and please treat them so. If the editors do not give a benefit of doubt, they will be killed.
At this state of domain knowledge, I do not expect that any of you are ready to confidently evaluate whether this picture is true, as it requires the entire landscape of psychology, cognitive science, neuroscience, computer science, electrical engineering and mathematics. All I am asking is: please let me know if you found any major flaws --- not simplifications but major flaws. Are they worth publishing? Please feel free to reply to all or reply to me only.
Also, if you are interested in undertaking this line of research, or would like to work on or have suggestions for improving the existing infrastructure, please send me an email so that we can discuss your ideas.
Enjoy your weekend.
Best regards,
-John
-------- Original Message --------
Subject: Re: A brain-mind picture: Your expert's vision, comments, and suggestions are needed
From: Paul Werbos <pwerbos@nsf.gov>
Date: Fri, 2 Apr 2010 09:04:12 -0400
Cc: suppressed
A dialogue of this distinguished group could possibly be very productive, but only if we do not try to put too much burden on each other. So I will just tell a simple story about one of the many points you raise..
On Apr 2, 2010, at 8:13 AM, taymore2002@aol.co.uk wrote:
> Marvin Minsky told me a cold fact: artificial neural networks cannot reason.
This sentence is meaningless without a definition of the term "artificial neural networks." Crudely -- to avoid wasting time -- let us say that it includes any implementation of any mathematical model of signalling and learning by a network of neurons, where the signalling between neurons follows the usual paths of
axons and dendrites (in forwards or backwards directions). (We usually use a broader definition, while people opposed to research in the area have sometimes tried to restrict it to something narrower.)
***IF** his brain is made up of such a network of neurons, then Minsky is saying that he cannot reason.
Karl Pribram has occasionally asserted that the brain really cannot be explained in such a way, that the essential signalling in the overall system involves some kind of field and quantum effects, generating a holographic mind which none of the family of models allowed by computational neuroscience today could possibly capture. And so, years ago, when NSF was planning for a new agency-wide initiative on Learning and Intelligent Systems, Karl was quite delighted when I told him: "the management says it is too narrow to talk about neural network models used both in neuroscience and in engineering. They say we need to make it more open." As he smiled a broad smiled and asked what more they had in mind, I sadly replied: "They said we need to make it open to the geology directorate."
The topic deserves a more complete statement, but I would not want to impose more on this entire group right now.
Best regards,
Paul
-------- Original Message --------
Date: Fri, 2 Apr 2010 10:59:42 -0400
Subject: Re: A brain-mind picture: Your expert's vision, comments, and suggestions are needed
From: Marvin Minsky <marvin.minsky@cba.mit.edu>
To: Paul Werbos <pwerbos@nsf.gov>
Cc: supressed
(Minsky's email was withdrawn per his request.)
-------- Original Message --------
Date: Fri, 02 Apr 2010 14:59:09 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Marvin Minsky <marvin.minsky@cba.mit.edu>
CC: suppressed
Subject: Re: A brain-mind picture: Your expert's vision, comments, and suggestions are needed
Dear Marvin:
Your formal version in a paper in the AI Magazine was something like artificial neural networks do not reason well. I attended a few of your interesting talks at MIT, where you stated an oral version "neural networks cannot reason." Regardless of the versions, I fully understood that you did not mean shallow ANDs, ORs, and NOTs. The point you made in those prior venues is valid.
Yes, one can always argue that he can build a very deep network using more of these ANDs, ORs, and NOTs. So what? Your criticism in your last email toward such deep networks is also valid, I think.
This is an issue that has been challenging for many researchers in Psychology, Cognitive Science, Neuroscience, Biology, Electrical Engineering and Mathematics to understand. A background in computer science helps.
Among many key points in brain-like reasoning by the new Epigenetic Developer (ED) (see file 2, name suppressed) are
(a) how to automatically find equivalent subsequences in a very long sequence of individual life, and
(b) how to do so without going back to the long sensorimotor life sequence itself because the raw sensory and motor signals are lost as soon as its time is passed.
The classic Finite State Machine (FSM) uses a human static design to specify such equivalences among many subsequences (e.g., "young cat" and "kitten" are equivalent in a long sensory sequence). Inspired by the brain anatomy, the ED showed that it can autonomously generate internal representations (network) to predict such many equivalences. Currently, the ED uses motor-imposed supervised learning. Of course, ED can use reinforcement learning and the new communicative learning modes.
I will take the liberty to post your email onto our IJCNN 2010 Panel Session web: http://www.cse.msu.edu/ei/IJCNN10panel/ until you tell me otherwise.
-John
-------- Original Message --------
To: suppressed
Subject: Re: A brain-mind picture: Your expert's vision, comments, and suggestions are needed
Date: Fri, 02 Apr 2010 08:13:38 -0400
From: John Taylor
Dear John,
I found your papers appealing but somehow missing the beat, especially the third. I attach two papers of mine, one summarising the EU GNOSYS project's creation of a reasoning robot and the second I have about to come out in Schizophrenia Bulletin. In particular in the latter I deal with consciousness and its neural creation. I hope these indicate my own approach: a functional one (different brain modules have different functions, given by genetic labels to provide diffeent guiding chemicals to generate different functionalities of modules by diffeent connectivities). I hope these aare of interest.
Best
John
-------- Original Message --------
Date: Fri, 02 Apr 2010 15:51:51 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: John Taylor
CC: suppressed
Subject: Re: A brain-mind picture: Your expert's vision, comments, and suggestions are needed
> different brain modules have different functions, given by genetic labels to provide different guiding chemicals to generate different functionalities of modules by diffeent connectivities
As far as I know from developmental biology and developmental neuroscience, genes regulate how each neuron moves and changes its shape (e.g., axon growth guided by neuronal activities via guiding chemicals as you said) and other characteristics, but genes do not rigidly specify functionalities of modules. Read Mriganka Sur's work about cross-modality plasticity (Sharma et al. Nature, 404, 841-847, 2000). He showed that when visual signals from the eyes wire into auditory cortex, the auditory cortex does visual information processing. Please read the second attachment to see how the functions of ED are generated using signals from sensors and motors. The ED model fits the evidence from Mriganka Sur's work.
-John
-------- Original Message --------
Subject: more on why I agree with what Marvin Minsky said to this list
From: Paul Werbos <pwerbos@nsf.gov>
Date: Fri, 2 Apr 2010 15:28:34 -0400
Cc: suppressed
To: Juyang Weng <weng@cse.msu.edu>
Please forgive any lack of clarity in what I sent out.
I strongly agree with the proposition suggested here by Marvin Minsky that recurrent (as opposed to loop free) networks are essential to anything like brain-like performance -- even mundane mouse-like performance -- in higher-order tasks.
Speech recognition is not the same as "reasoning," but of course there is a relation. Many years ago, work by Principle using time-lagged recurrent networks did an order of magnitude better than anything using feedforward neural networks in phoneme recognition, in serious tests performed by folks in some other agencies.
(Actually, I think it was an order of magnitude better than anything else at all...) Likewise, I used to cite a Japanese guy form that time who found that RECURRENT networks could solve grammar parsing kinds of problems that Fodor talked about, that could not be handled in any other way. The connectedness example from the book Perceptrons is important, and used and cited in that 2008 paper I did with Kozma and Ilin.
One caveat:
This past year both Yann LeCun and Andrew Ng have reported some breakthrough results in object recognition and in phoneme recognition. They used standard challenge benchmark datasets, so perhaps I need to cite them in future instead of others. I have not yet seen the relevant papers, but I don't recall any recurrence in what I have seen so far regarding their architecture (unless you count backpropagation in the learning process, which is not sufficient by itself for the kind of results Kozma, Ilin and I had). But there are more fundamental reasons to believe the point that Marvin has said on this occasion.
As for the past... I strongly doubt that any one of us could survive total scrutiny of the accuracy of everything we have ever said. I hope we can all move on to cooperate more effectively to support the kinds of larger goals that the NFS COPNprogram was going for (again, expanded in its mathematics component).
Best of luck,
Paul
-------- Original Message --------
Date: Fri, 02 Apr 2010 16:58:45 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Paul Werbos <pwerbos@nsf.gov>
CC: suppressed
Subject: Re: more on why I agree with what Marvin Minsky said to this list
Paul:
A major limitation of existing artificial neural networks, feedforward and recurrent, is that they cannot go beyond recognition or regression. Dealing with time warping like a speech recognition system is not sufficient for the brain.
You did not get the points that Marvin Minsky was talking about. He was right! Please read my reply to Marvin. I guess that what is missing between Marvin Minsky and you is the automata theory. I guess that your background is from electrical engineering and artificial neural networks. Please read or review the theory of automata and formal language theory.
Many of our neural network panelists have turn their mind off using their pre-assumed neural network models for many years ... If they do not understand automata theory, they will never understand the brain.
That is why I afraid that the brain-mind picture that I presented in the two files will be very difficult for psychologists and neuroscientists to confidently appreciate. Many of them do not understand how an FSM operates. The same is true for an electrical engineers and professors if they do not have solid training of automata and formal language theory.
In the new ED theory (attachment 2), the brain is not only a finite state machine (FSM), but a self-generative statistically optimal FSM, much more powerful than any HMM or POMDP (partially observable Markov Decision Process) that you know. That is, ED is epigenetic: it grows using sensory signals and motor signals regulated by the "genome program" --- developmental program.
The attachment 3 explains a richer picture of the brain-mind with 5 chunks, development, architecture, area, space and time. Missing any chunk of the five will be far from the brain. For example, it explains how the brain deals with time beyond 50ms (explained by Spike Time Dependent Plasticity STDP).
Could you please read the one-journal-page attachment 2? It goes far beyond the recognition or regression that you are talking about. It abstracts and reason on the fly, automatically finding many subsequences that are motor-equivalent.
Please, please, please read them. Let me know which part is not clear.
-John
-------- Original Message --------
Date: Fri, 02 Apr 2010 17:45:37 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Minoru Asada <asada@ams.eng.osaka-u.ac.jp>
Subject: Between Bottom-Up and Top-Down What Is “the Much in-between”?
From Minoru Asada:
> Is “Between Bottom-Up and Top-Down What Is “the Much in-between”?” a right question?
> “Connectionist approaches are bottom-up (e.g., from pixels)”
> --- It seems, but what is the most bottom?
> “symbolic approaches are top-down (e.g., from abstract concepts)”
> --- It seems, but where can abstract concepts be given from?
> The right question seems “How embedded structures and social interactions enables cognitive development of humans from viewpoints of different levels and timescales?”
Minoru:
The questions we raised are for understanding the development of the brain-mind, regulated by the genome.
For two main reasons, AMD should not "embed structures" that are task specific into the machine's brain:
(1) Such embedded structures are brittle, as they are static, assumed. During deployment, such assumptions are typically violated. Then the system breaks.
(2) The brain is "skull-closed" during development. It has only two terminals to interact with its environment during its autonomous
development: sensors (rods and cones for eyes and pixels for cameras) and effectors (muscles and glands). Therefore, the brain is very plastic. It can general distributed internal representations from sensor and motor experience.
As you are asking "the most bottom", I guess that you mean the physical world. Please consider the following question: Does the baby's brain have an explicit model about the baby delivery room before the birth time?
As to the question about top-down, please read the postings at http://www.cse.msu.edu/ei/IJCNN10panel/. Some additional postings (e.g., from Marvin Minsky) are from separate emails.
-John
-------- Original Message --------
To: pwerbos@nsf.gov
Subject: Re: A brain-mind picture: Your expert's vision, comments, and suggestions are needed
Date: Fri, 02 Apr 2010 09:40:12 -0400
From: John Taylor
Cc: suppressed
Dear Paul and colleagues,
I am afraid it would be incorrect to ascribe to my email that it wrote that 'neural networks cannot reason'. Indeed the GONSYS review paper (and several other papers from our group) show that a suitably defined set of neural architectures can reason (at a non-linguistic level), such as solving the 'two sticks''.problem, a task soluble by New Caledonian crows (Betty et al).
I have great respect for Karl Pribram who I think of as a great man but I cannot accept his comments either. Nor I am afraid the gist of your argument - as I take it you are saying that some sort of quantum effect is needed to get neural networks on a firm basis. Shades of Sir John Eccles and Sir Roger Penrose, to add further great names to the mix. But there is still a lot of leeway in the present work on looking at the flow of information in attention control paradigms in the brain (see Desimone et al Science paper, 2009, and Steve Bressler et al,J Nsci 2008). And they are only just scratching the surface of infomation flow in the brain! A long way to go before we should give up on the neural networks of the brain and turn to the modern alchemy of quantum mechanics.
Best
John
-------- Original Message --------
Date: Fri, 02 Apr 2010 20:46:04 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: John Taylor
CC: suppressed
Subject: The 5 chunks of our brain-mind model
Dear John,
I would like to briefly discuss some basic points of ED in file 2 and the more complete account of the brain-mind picture in file 3. I hope that this is useful for those from natural intelligence who probably will not be able to fully understand the material in it.
While doing it, I guess that it might be good to raise the basic differences between your GNOSYS described in your review paper (I call it G paper below) and our brain-mind model in file 3. I hope that it can motivate you a little to spend your valuable time to read file 3.
Our brain-mind model consists of 5 chunks, which are very concisely concentrated into the Epigenetic Developer (ED) in file 2:
Chunk 1: Development. According to the the genomic equivalence principle (discussed in file 3), every neuron is totally autonomous during development, as our ED systematically modeled. Therefore, hardly any brain module has pre-specified function. This is consistent with Mriganka Sur's discovery. All the functions of all brain areas are emergent from experience.
However, in Fig. 2 of the G paper, all the modules have a handcrafted function. For example, a module is called "concept" and another is called "reasoning". Such clear-cut functions seem inconsistent with the the genomic equivalence principle.
You might want to re-consider this scheme. The Genomic equivalence principle is really powerful for us to understand how functions emerge from the developmental brain through experiences.
Chunk 2: Architecture. According to rich studies of neuro-anatomical studies (e.g., the detailed tables in Felleman & Van Essen 1991), if an area A sends signals to B, it must receive signals from B. There are few exceptions (e.g., between retina and LGN). All the wires between brain areas and within areas are functions of experiences.
However, in Fig. 4 of the G paper, from LGN up along the ventral stream, and from LGN up along the dorsal stream, the connections are all one way (up) only. Such circuits are not able to deal with fine-grained hierarchical top-down visual attention, and cannot deal with complex backgrounds. I guess from Figs. 16-17 that your project assumed objects of interests have special colors.
Chunk 3: Area. Our file 3 presents a scheme of laminar cortex, as a general-purpose developmental block of the brain. It presents a detailed wiring diagram of the general-purpose unit. Our LCA model explains how such a block develop best features. They are dually optimal: spatially and temporally.
The G. paper did not have a general-purpose model of cortical area. In fact, this is one of the basic problem of many existing artificial neural networks. Each layer must be learned using different learning expressions. I hope that our LCA model is useful to your future work.
Chunk 4: Space. The most basic and amazing function of spatial processing of the brain is to dynamically pick up only a small part of information from the sensors and effectors to reason. Our model in file 3 is the first general-purpose model for this much overlooked function of brain's spatial processing. I would like to say that natural intelligence is displayed through attention.
The architecture in Fig. 4 of G. paper does not allow such a general purpose spatial attention. For example, it used mostly one-way connections.
Chunk 5: Time. The ED in file 2 presents the first general-purpose scheme about how a developmental brain deals with general purpose time without dedicated temporal component. This miracle is required by the genomic equivalence principle.
On page 8, the G. paper uses time delays ST(t-4), ST(t-3), ... , ST(t) to deal with a span of 5 time frames. This does not appear to be how the brain deals with time. This way cannot deal with the general time warping that Paul Werbos talked about when he mentioned HMM. The brain must deal with temporal context of any desirable length, not just a fixed 5 frames. However, if the temporal length increases, the number of possible trajectories exponentially explodes, but ED avoids this big problem.
There are many other aspects that the biological brain does beautifully and we modeled in file 3. I hope that the above five chunks are useful to you.
-John
-------- Original Message --------
Date: Sat, 3 Apr 2010 12:48:17 -0400
Subject: Re: A brain-mind picture: Your expert's vision, comments, and suggestions are needed
From: Marvin Minsky <marvin.minsky@cba.mit.edu>
To: Juyang Weng <weng@cse.msu.edu>
(Minsky's email was withdrawn per his request.
Instead I quote the exact text from: M. Minsky, Logical versus analogical or symbolic versus connectionist or neat versus scruffy, AI Magazine, 12(2), 34-5, 1991:
Page 36, the 1st paragraph: "Our purely numeric connectionist networks are inherently deficient in abilities to reason well;"
Page 47, the 1st paragraph: "In section 6.6 of Perceptrons (Minsky and Papert 1988), we see that we must be prepared to find that even after training a certain
network to recognize a certain type of pattern, we might find it unable to recognize this same pattern when embedded in a more complicated context or environment. (Some reviewers have objected that our proofs of this fact applied only to simple three-layer networks; however, most of these theorems are much more general, as these critics might see if they’d take the time to extend those proofs.)
-John Weng
)
-------- Original Message --------
Date: Sat, 03 Apr 2010 14:09:33 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Marvin Minsky <marvin.minsky@cba.mit.edu>
CC: suppressed
Subject: Re: A brain-mind picture: Your expert's vision, comments, and suggestions are needed
Dear Marvin:
This is a free country and we enjoy full academic freedom. You as a respected scholar and a leader in AI, should you dodge on the well known problem of prior artificial neural networks that you have honestly raised? This prior problem is not your fault. You are helping others by pointing it out early. I think that this is your way of contribution to the brain science, regardless whether others think of it the same way. In fact, your perspectives and repeated criticisms have prompted me to successfully solve this problem using my over 20 years of time.
My file 2 shows that this network-reasoning problem has been successfully solved by the new Epigenetic Developer (ED). The brain inspired network ED can reason as a general purpose framework of FSM (and its variants HMM, POMDP, Bayesian nets, belief nets, etc.), on which many AI knowledge-base systems are based. Furthermore, the ED is emergent, fully autonomous in wiring up and adapting its brain-inspired internal representation (i.e., skull-closed), instead of a handcrafted static FSM which does not have any such internal representation.
Thank you for checking it out. I still have not heard any comment that points out any major flaw of ED in file 2. This email list has included well matched, distinguished researchers in psychology, cognitive science, neuroscience, computer science, electrical engineering and mathematics.
-John
-------- Original Message --------
Date: Sat, 03 Apr 2010 15:31:52 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Marvin Minsky and others (suppressed)
Subject: Your exact wording
Dear Marvin:
Your email has been withdrawn from the web per your request. My sincere apology.
I think your email clarification is similar to the following exact text I extracted from M. Minsky, Logical versus analogical or symbolic versus connectionist or neat versus scruffy, AI Magazine, 12(2), 34-5, 1991:
========
Page 36, the 1st paragraph: "Our purely numeric connectionist networks are inherently deficient in abilities to reason well;"
Page 47, the 1st paragraph: "In section 6.6 of Perceptrons (Minsky and Papert 1988), we see that we must be prepared to find that even after training a certain network to recognize a certain type of pattern, we might find it unable to recognize this same pattern when embedded in a more complicated context or environment. (Some reviewers have objected that our proofs of this fact applied only to simple three-layer networks; however, most of these theorems are much more general, as these critics might see if they’d take the time to extend those proofs."
========
By the way, the brain-inspired framework ED (file 2) has shown that a three-layer ED is indeed able to "recognize this same pattern when embedded in a more complicated context or environment." However, it requires some experience for different locations, sizes, viewing angles, etc., since this seems also largely what the brain needs, as demonstrated by the studies cited by our Cresceptron paper in IJCV 25(2) 109-143, 1997. However, the brain (and ED) is not a rote learner! The innervation of more ED areas into an existing ED network enables increasingly more complicated thinking processes which infer for new locations, sizes, viewing angles that ED has not actually observed, using the new ED framework of top-down connections. This capability of ED to learn to autonomously think and discover is yet to be further
explored and demonstrated.
-John
-------- Original Message --------
Date: Mon, 05 Apr 2010 13:39:02 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: suppressed
Subject: Reply to Pinaki Mazumder at a brain-mind solution
Pinaki Mazumder wrote: <suppressed>
Pinaki:
Thank you for asking. This is probably one of the largest scientific and technology news in your life time, at least I guess it is because of your area of research in VLSI circuits. If you are interested in this brain-mind solution, read the files 1, 2, and 3 that I attached in the first email.
You probably can understand it if your mind gets into it because you are analytical. But I am afraid that currently you cannot adequately appreciate it, because you need sufficient width and depth of the entire landscape (psychology, cognitive science, neuroscience including related biology, computer science, electronic engineering which you have, and mathematics) to do it.
STDP is an important but smaller story for the brain because it is about a single neuron. Mriganka Sur said in a seminar at MIT that the big stories with the brain are about many neurons. STDP seems a special form of Hebbian learning according to Mu-ming Pu at Berkeley. ED and WWN model STDP as digital spikes, useful when the network update rate has reached around 1000hz for better temporal precision.
If you did not book mark it, please read the emails in http://www.cse.msu.edu/ei/IJCNN10panel/.
As you send this email to me only, I do not post your email till you tell me to.
-------- Original Message --------
Date: Sun, 4 Apr 2010 11:23:41 -0400
Subject: Re: Your exact wording
From: James Albus <james.albus@gmail.com>
To: Juyang Weng <weng@cse.msu.edu>
Cc: suppressed
John,
The problem here is that segmentation of an object from the background is a prerequisite to reasoning. A neural network cannot be trained to recognize an object against all possible backgrounds. Until the object is segmented from the background, all other processes such as computation of attributes, classification, recognition, and cognitive reasoning cannot be robust.
Jim Albus
-------- Original Message --------
Date: Mon, 05 Apr 2010 13:14:07 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: James Albus <james.albus@gmail.com>
CC: suppressed
Subject: Dynamic receptive fields
Jim:
Did you read all the three files?
I am glad to hear your comment. Your concern of background is valid. The files 2 and 3 spent much space to address the background issue.
Files 2 and 3 were the first one, as far as I know, that addressed this general-purpose background issue. Each neuron has a default receptive field, which further fine tunes through experience. Our ED model (file 2) and the rich version WWN (file 3, e.g., figure 9) indicate that the receptive field is not necessarily a connected static region or static. In fact, in general, a receptive field is disconnected and is dynamic, all determined by top-down attention at each time instant. All these aspects of a receptive field
are shaped from experience.
You wrote "A neural network cannot be trained to recognize an object against all possible backgrounds." This is exactly not what ED and WWN do. In WWN, each receptive field has a limited scope and is further fine tuned and dynamically shaped by top-down attention.
One might doubt whether the brain has so many receptive fields for all the possible retinal locations and all the practical sizes. Tsotsos has proved that finding the correct receptive field is NP-complete. However, he only considered static receptive fields.
The brain is smarter. How? A major key is: "dynamic".
For example, according to the WWN model, every motor neuron in the brain has the entire retina as its receptive field. This is because virtually every "pixel" in the retina can totally determine the neuron's action. Which pixel depends on the current attention (e.g., when your brain is examining the color of a particular pixel). How can the brain develop this capability of dynamic receptive fields? ED and WWN indicate that the receptive field of every their motor neuron is highly dynamic this way. Currently, with only one default size of receptive field, every motor neuron of WWN can already attend to every pixel location! If WWN has more areas, it can attend individual areas as small as every "pixel" or as large as the entire image.
If you have not, please read ED in file 2 and WWN in file 3 to see how.
-John
-------- Original Message --------
Date: Thu, 08 Apr 2010 15:26:30 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: suppressed
Subject: Re: Dynamic receptive fields
> Dear John,
>
> Please point me to a technical description of your proposed solution
> of recognizing complex objects in naturalistic backgrounds, including
> benchmark studies.
>
> Also, please do NOT copoy your reply to the entire email distribution
> list.
>
> Thanks!
>
> xyz
> Dear xyz:
>
> The file 2 (Title: Epigenetic Developer) has addressed this issue,
> although its size is much condensed.
>
> See Fig. 1. It has three parts (a), (b), and (c). In (c), you can
> see that each example has two learned objects in complex, natural
> backgrounds. Depending on the intent (requested by the operator or
> self-generated), WWN reports either the type or location.
>
> More detailed bench marks are in Fig. 2 (c). The recognition rate
> was about 95% and location error is under 1.3 pixels.
>
> -John
Dear xyz, your name was removed so as to keep your ID private. The
text material is useful for others. This is not the same as Minsky's
case, as the material he discussed is directly related to his ID. Your
text material is not related to your ID, very generic.
-John
-------- Original Message --------
From: Paul Werbos <pwerbos@nsf.gov>
Date: Mon, 5 Apr 2010 13:45:43 -0400
Cc: suppressed
To: Juyang Weng <weng@cse.msu.edu>
On Apr 5, 2010, at 1:14 PM, Juyang Weng wrote:
>
>
> You wrote "A neural network cannot be trained to recognize an object against all possible backgrounds." This is exactly not what ED and WWN do. In WWN, each receptive field has a limited scope and is further fine tuned and dynamically
> shaped by top-down attention.
As I understand it, this is what the neural networks of Ng and of LeCun have done this past year -- better than any of the classical types of systems, reasoning-based or other -- demonstrated on such a challenge database for object recognition used by many competitors. DARPA believes it enough to have given them both major follow-on grants.
The neural networks in the brain of a reptile are not enough to "reason" (as "reasoning" is usually understood in the English language) but they are enough to make out objects in a visual field. And of course birds do even better, in a rough sort of way.
The classic paper of Albus from 1991, An Outline of Intelligence does raise very important challenges to folks modeling the neurons of the brain or trying to build mathematical designs capable of either modeling or replicating their capabilities. (It's the first citation in my paper on this subject in Neural Networks this past year.) Trivial types of neural network model cannot meet those challenges, but they can be met. I would refer back to my first email to this group... which I will not repeat.
But I can't spend more time on this now, because there are major deadlines here....
Best of luck,
Paul
-------- Original Message --------
Date: Mon, 05 Apr 2010 15:06:11 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Paul Werbos <pwerbos@nsf.gov>
CC: suppressed
Subject: Re: Dynamic receptive fields
Paul:
Thank you. I am aware of Ng and LeCun's recent work. Their past work used error back-propagation. Their recent work also used two-way connections. Their local filters look at ICA filters. Their work is similar to Geoffrey E. Hinton's Deep Belief Networks, which I cited in file 3.
The reasoning Marvin Minsky was talking about is the goal directed search, where the goal may change all the time. In symbolic world, this is not very difficult. For networks that perceive the real world with backgrounds, there has been no solution till our ED (file 2) and WWN (file 3).
Ng, LeCun, and Hinton's networks do not deal with complex backgrounds. Their networks use unsupervised learning, thus, cannot do goal directed search in endless reasoning.
The major missing part of their work is the rule of motor areas for reasoning and the rule of motor area to pick up foreground from complex backgrounds. This is exactly what Marvin Minsky criticized to be missing and what our ED and WWN have demonstrated. The convolution of the networks of LeCun and Ng is not what the brain does, but this is minor regard to the above major misses.
Albus' work in 1991 and 2010 (I cited) used open network. This is a major reason that it cannot reason endlessly as Marvin Minsky criticized to be missing.
They are all good work. However, we are talking about a brain-mind picture, which must reason endlessly in complex backgrounds. We are not taking about just neural networks.
By the way, HMM in speech recognition typically does not use FSM in its full form (e.g., left to right version only).
I hope that the above clarification is useful.
-John
-------- Original Message --------
Subject: Re: Dynamic receptive fields
From: Paul Werbos <pwerbos@nsf.gov>
Date: Mon, 5 Apr 2010 15:58:47 -0400
Cc: suppressed
To: Juyang Weng <weng@cse.msu.edu>
================================
I do not think one could construct or explain a brain so advanced as a lamprey without always being clear about the distinction between "reasoning," between recognizing objects against a complex background, and goal-directed behavior. (And, yes, they have connections, just as Republicans and Democrats do in Washington, but I really wonder about the folks who ask us to explain how aplysia does modus ponens.)
I do not claim that the recent work of LeCun and Ng is as advanced, say, as the brain of a reptile, in processing images. But it already does well on the task of recognizing objects against a realistic, challenging testbed. "More sophisticated" methods or designs which cannot do the same ... have no justification for asserting that this empirical domain provides support for preferring their theories. This domain provides no evidence to support the claim that "neural networks" as I defined the term in my first reply cannot.. as you say... recognize objects against a complex background.
======
I am amazed that, after this discussion, anyone would still cite Marvin Minsky as endorsing a position he does not endorse.
If human beings were truly born reasoning beings, I do not think they would make that kind of ellision so easily or so persistently, or gloss over the rather crucial issue of what definitions we assume for words when making strong assertions. In my view, the problem of explaining human nonreasoning is just as important as explaining human reasoning.... and not so realistic unless one understands, say, the higher intelligence of the mouse much better than folks generally do today.
Best of luck,
Paul
-------- Original Message --------
Date: Mon, 05 Apr 2010 17:33:26 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Paul Werbos <pwerbos@nsf.gov>
CC: suppressed
Subject: Let me try an example
Paul:
Sorry, that is why I am scared! Cannot explain to everybody!
Let me still try through email. I hope that nobody will regret his time of reading emails if he understands this brain issue. But please do not think that what I write below is sufficient, as the brain-mind picture needs 5 chunks. This is only a simple example.
Why is a recognizer not enough for the brain?
The brain needs to come up with a gaol or intent all the time and reason based on the intent and environmental evidence. Here is an example:
Time (discrete): t1, t2, t3, t4,
Eye sees: image1,image2,image3,image4, ...
Obj1 position: pos1, pos2, pos3, pos4, ...
Obj2 position: pos3, pos1, pos2, pos1, ...
Background changes: B1, B2, B3, B4, ...
Action from network: A1, A2, A3, A4, ...
(1) For a recognizer, action-i is the label of an object in image i. However, which one, obj1 or obj2? This problem of recognition from unknown background had no general purpose solution in the past.
(2) Thus, the brain needs self-generated intent: How does the network/brain generate a goal at every time ti? For example, obj1 or obj2, pos2 or pos3? Where does it generate intent?
(3) Goal directed reasoning: How does the network/brain find location of obj1 if the intent is obj1, or tell the object type at pos2 if the intent is pos2? One can always write a special purpose program given above two tasks. But no, you do not know the above two tasks till you have finished your learning program.
(4) What about a general intent? No, you do not know even what the robot will learn about in the future.
(5) What about giving the reasoning solution for such a general intent? No, you do not know even the nature of its future intent.
(6) The brain must be skull-closed: The teacher cannot open the brain to directly intervene its full autonomy of internal self-organization. The teach can interact with the motor and the eye only.
Among many other things, the above 6 problems have been resolved by a single unified brain-like ED and its more extensive version WWN.
An example of "the other things" is time: The reasoning result at time ti may need information in
an unknown number of time frames in the past. This temporal dependency is implicit, not known to your learning program.
-John
-------- Original Message --------
Subject: Re: Let me try an example
From: Paul Werbos <pwerbos@nsf.gov>
In-Reply-To: <4BBA5726.7090001@cse.msu.edu>
Date: Wed, 7 Apr 2010 11:27:51 -0400
Cc: suppressed
To: Juyang Weng <weng@cse.msu.edu>
You are asking important questions, and deserve some answer, but I have a prior duty to answer questions which came to here through NSF
proposals. Perhaps I will have more time later this week.
-------- Original Message --------
Date: Wed, 07 Apr 2010 22:30:36 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Paul Werbos <pwerbos@nsf.gov>
CC: suppressed
Subject: Re: Let me try an example
No problem. By the way, they are not just questions. They are what our Epigenetic Developer (ED) and Where-What Networks (WWNs) have demonstrated (files 2 and 3).
As there is nothing in their developmental programs that specifies the meaning of motor response, in principle such networks can be taught with, and can self-generate, other types of intent.
As the models are rooted in brain's anatomy, they model brain's ways of reasoning that have been missing in the prior artificial neural networks that Marvin Minsky criticized. According to ED and WWN, top-down attention is a major way for the brain to reason autonomously in space and time.
However, all the five chunks (development, architecture, area, space and time) solved each chunk's well known bottleneck problems before ED and WWN have finally showed how top-down attention reasons for some small scale, realistic real-world data. In other words, a developmental program for general purpose "brains" has been fully specified at the algorithm and software levels.
-John
-------- Original Message --------
From: Paul Werbos <pwerbos@nsf.gov>
Date: Wed, 7 Apr 2010 15:31:29 -0400
Cc: suppressed
To: Juyang Weng <weng@cse.msu.edu>
On Apr 5, 2010, at 5:33 PM, Juyang Weng wrote:
>
> Why is a recognizer not enough for the brain?
I never said that a whole brain could be just a recognizer. In general, all our communities need to be careful about the claims
we make about other people's positions.
From IJCNNs, you know I have said MANY, MANY times that the brain as WHOLE SYSTEM is essentially an intelligent controller --
a system which maps from sensory inputs to "squeezing and squirting," with a whole lot of memory and other important stuff in-between.
Nothing I have said here contradicts that basic position.
You wrote:
Among many other things, the above 6 problems have been resolved by a single unified brain-like ED and its more extensive version WWN.
It seems that the unique advantages of your particular model are what you really want us to discuss.
Of course, one way to arrange that is to submit it as a proposal to somewhere at NSF (or multiple places).
That ensures far more attention than you could legitimately expect from this kind of email discussion.
I plan to sign off at this point.
-------- Original Message --------
Date: Mon, 05 Apr 2010 14:15:44 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: asada <asada@ams.eng.osaka-u.ac.jp>
CC: acangelosi@plymouth.ac.uk, nkasabov@aut.ac.nz, pasa@dist.unige.it, ASIM.ROY@asu.edu, rsun@rpi.edu, taymore2002@aol.co.uk, pwerbos@nsf.gov, parena@diees.unict.it, nsrinivasa@hrl.com, starzykj@gmail.com
Subject: Re: Between Bottom-Up and Top-Down What Is “the Much in-between”?
Minoru:
The goal of a developmental program is to simulate the functions of the genome.
Please consider the important issues raised for the Panel. You wrote "modified Hebbian learning and SOM both of which are main learning methods we have been using. But, we do not intend these are unique methods to utilise."
Why not? Are there any major problems in "modified Hebbian learning and SOM" in terms of their architecture and representations?
Please trust me, these questions will automatically and systematically take into account the body issue and social interaction issues you are interested in, but autonomously instead of handcrafted.
-John
-------- Original Message --------
Date: Tue, 06 Apr 2010 08:32:00 +0900 (JST)
To: weng@cse.msu.edu
Cc: acangelosi@plymouth.ac.uk, nkasabov@aut.ac.nz, pasa@dist.unige.it, ASIM.ROY@asu.edu, rsun@rpi.edu, taymore2002@aol.co.uk, pwerbos@nsf.gov, parena@diees.unict.it, nsrinivasa@hrl.com, starzykj@gmail.com, asada@ams.eng.osaka-u.ac.jp
Subject: Re: Between Bottom-Up and Top-Down What Is “the Much in-between”?
From: asada <asada@ams.eng.osaka-u.ac.jp>
> The goal of a developmental program is to simulate the functions of
> the genome.
This statement seems too strong. It seems that "genome" controls everything. The genome coding and the environmental issues complicatedly interact with each other in different levels. Therefore, only considering the algorithms lacks something on another side.
> Are there any major problems in "modified Hebbian learning
> and SOM" in terms of their architecture and representations?
Yes, there are some issues to design, of course. But, it depends on several aspects such as the task and/or the age of development to consider.
> Please trust me, these questions will automatically and systematically
> take into account the body issue and social interaction issues you are
Of course, but seems implicitly. I'd like to make them explicit from a viewpoint of design theory.
> interested in, but autonomously instead of handcrafted.
One of the issues is to what extent given and from where learning/development. Embryo, fetus, newborn baby, infant, toddler, child, juvenile, and, and, ...
Depending on the age, the issue changes, and the structure of the development is so complicated. Therefore, focusing on only the algorithm seems missing something.
Anyhow, I just made a question to the question. A different view (another extreme?) might be OK to make the panel more exciting:-)
Thanks lots.
-------- Original Message --------
Date: Mon, 05 Apr 2010 23:09:15 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: asada <asada@ams.eng.osaka-u.ac.jp>
CC: acangelosi@plymouth.ac.uk, nkasabov@aut.ac.nz, pasa@dist.unige.it, ASIM.ROY@asu.edu, rsun@rpi.edu, taymore2002@aol.co.uk, pwerbos@nsf.gov, parena@diees.unict.it, nsrinivasa@hrl.com, starzykj@gmail.com
Subject: Re: Between Bottom-Up and Top-Down What Is “the Much in-between”?
Minoru:
This issue was discussed repeatedly during Workshop on Development and Learning. "Genome" does not control everything. It only "modulates" the process
of development. This is a term from biology. The task of simulating genome program is to investigate how genome and environment interact to generate phenotype (human body, brain, mind, behaviors, etc.).
> Yes, there are some issues to design, of course. But, it depends on several aspects such as the task.
It is true that genome enables the emergence of some inborn reflexes, but genome seems not task specific. For example, the genome does not "know" that the child will be an engineer or artist when a life begins during conception.
You mentioned the age. Yes, age is taken care of by the genome, not rigidly, but taken into account.
> Embryo, fetus, newborn baby, infant, toddler, child, juvenile, and, and, ...
When a science is young, researchers can only talk about phenomena. We try to advance, so that we investigate underlying principles based on phenomena, such as the developmental mechanisms for the brain.
Are you interested in the issue of internal representation of the brain? This is about what is in the brain from interaction between the genome and the environment (including body, social interactions, age, etc.)
-John
-------- Original Message --------
From: Juyang Weng [weng@cse.msu.edu]
Sent: Wednesday, April 07, 2010 6:37 PM
To: Srinivasa, Narayan
Subject: Re: Between Bottom-Up and Top-Down What Is "the Much in-between"?
Narayan:
Thanks.
By the way, how do you think about the following view?
The brain is composed of special areas. For example, an area has concepts and another area has values.
-John
-------- Original Message --------
From: Srinivasa, Narayan <nsrinivasa@hrl.com>
To: Juyang Weng <weng@cse.msu.edu>
Date: Wed, 7 Apr 2010 20:14:39 -0700
Subject: RE: Between Bottom-Up and Top-Down What Is "the Much in-between"?
Hi John,
My point of view is that the brain may appear specialized due sensor specific pathways but the fact that they are connected in more ways than just the feedforward paths that you normally consider, they fully affect each other. In fact there are more feedback/reentrant pathways than feedforward. So, even if there are special areas (be it for value or concepts), they are still affected by the activity in other areas. So, binding/association in space and time is unavoidable and persistent.
Best,
Narayan
-------- Original Message --------
From: - Thu Apr 08 15:15:35 2010
Date: Thu, 08 Apr 2010 14:55:56 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Srinivasa, Narayan <nsrinivasa@hrl.com>, Minoru Asada <asada@ams.eng.osaka-u.ac.jp>, Angelo Cangelosi <acangelosi@plymouth.ac.uk>, Nik Kasabov <nkasabov@aut.ac.nz>, Giorgio Metta <pasa@dist.unige.it>, Asim Roy <ASIM.ROY@asu.edu>, Ron Sun <rsun@rpi.edu>, John Taylor <taymore2002@aol.co.uk>, Juyang (John) Weng <weng@cse.msu.edu>, Paul Werbos <pwerbos@nsf.gov>, Paolo Arena <parena@diees.unict.it>, Narayan Srinivasa <nsrinivasa@hrl.com>, Janusz Starzyk <starzykj@gmail.com>
Subject: Re: Between Bottom-Up and Top-Down What Is "the Much in-between"?
Srinivasa,
I agree with you. As what you wrote is consistent with our panel's discussion, I also provide a CC to our panel.
-John
-------- Original Message --------
Subject: what neural networks can do
From: Paul Werbos <pwerbos@nsf.gov>
Cc: Minoru Asada <asada@ams.eng.osaka-u.ac.jp>, Angelo Cangelosi <acangelosi@plymouth.ac.uk>, Nik Kasabov <nkasabov@aut.ac.nz>, Giorgio Metta <pasa@dist.unige.it>, Asim Roy <ASIM.ROY@asu.edu>, Ron Sun <rsun@rpi.edu>, John Taylor <taymore2002@aol.co.uk>, Paolo Arena <parena@diees.unict.it>, Narayan Srinivasa <nsrinivasa@hrl.com>, Janusz Starzyk <starzykj@gmail.com>
Message-Id: <C01EA1E2-4991-41D1-B890-9D22E8036705@nsf.gov>
References: <4BB797A8.3010605@cse.msu.edu> <t2qa346ad7c1004040823m50129d74y960448db7d4c3c50@mail.gmail.com> <4BBA1A5F.4070503@cse.msu.edu> <906C658B-02DE-4410-8C2D-1725C16A72A5@nsf.gov> <4BBA34A3.4080902@cse.msu.edu> <A45A01CA-0527-42B1-B880-D1F6737856AE@nsf.gov> <4BBA5726.7090001@cse.msu.edu> <4BBE3933.2010509@cse.msu.edu>
To: Juyang Weng <weng@cse.msu.edu>
John has raised some important questions which need to be looked at more precisely, without associating them too much with personalities.
Crudely, one set of questions can be expressed: "Can a neural network perform object recognition against a background, or effective goal directed behavior, or other things which we know even a mouse can learn to do?"
We cannot really discuss this in a rational way, without selecting WHICH CONCEPT or definition of "neural network" we assume when asking the question.
In my first email, I basically defined a neural network as a system made up of the following kinds of objects:
M types of "neuron", plus the sensor type and the motor type (call them types M+1 and M+2);
a collection of n-sub-i neurons for each type i from 1 to M+2;
(let us assume an index number j to identify the neurons);
for each neuron j, a list of m-sub-j "neighbors";
for each neuron j, a vector of current inputs and a vector of current outputs at each time t;
for each neuron j, some kind of large vector or structure X-sub-j(t) describing its internal state;
for each type of neuron, a dynamic law (possibly stochastic) which specifies X-sub-j(t+delta-t)
as a function of X-sub-j(t), inputs-to-j(t), and maybe event outputs-from-j(t);
likewise, a dynamical law to specify its outputs.
Plus perhaps a short vector of global "chemical" variables, not indexed by j, but indexed by t.
We are looking for a neural network which can do far more than any one of the neurons
which it is made up of, and a certain degree of sparseness in the connectivity, and where connectivity is
present only for neurons connected by axons or dendrites.
This is probably formal enough for our purposes here and now. Let's call this a "normal general neural network (NGNN)."
Unlike certain other people I know, I would claim that NGNNs can be designed (with appropriate dynamical laws) which can indeed learn to recognize objects against complicated backgrounds, and perform all the higher-order kinds of things that a single mouse brain could ever evince in a laboratory (not counting claims of paranormal abilities in mice). Maybe Karl Pribram would disagree with me on this; maybe not. I would even claim that NGNNs could learn to do "reasoning" as Marvin Minsky was referring to, or as George Bernard Shaw was implicitly discussing in Back to Methusaleh.
However, there is a special case of NGNN, a subset, which Pribram has called the "conventional neural dogma" (CND). In CND, the output vector consists of just a single number sent forwards along the axon. The neurons (leaving aside hardwired motor pool neurons and such) are all governed by dynamics which are invariant with respect to time translation, and local in time, and subject to learning. In essence, the dynamics are governed either by well-posed ordinary differential equations or by a slight generalization, allowing for spikes.
Pribram and I agreed on our very first meeting that CND cannot even meet the lowly standards of functional capability we see in simple engineering systems, let alone the more complex things that mice brains can learn to do.
The most obvious thing they are missing, in my view, is not some sort of quantum field effect, but the lowly CLOCK. Of course, one can fake a clock (or explain it) by inserting nonlearning, synchronized oscillator cells, which modulate all the "real" neurons. That's basically how the machinery of the brain handles this. Back in the 1970's people like Purpura and Scheibel and Scheibel sketched out the anatomy of how the "nonspecific" timing signals from the nonspecific thalamus modulate the giant pyramid cells of the cerebral cortex. (See the old Rockefeller books, numbers one and two, edited by F. Schmitt.) Foote had a more general review of this in the 1990's, in Annual Reviews of Neuroscience, and LLinas has gone on to show how the main clocks in the brain are synchronized with millisecond precision. There are some functional tasks in engineering which can be done in continuous time (as in some of the recent work of Frank Lewis), but even there sampled time and clocks are a necessary part of the system, to make it work.
Beyond that, I would put forward the hypothesis that local back-flowing signals are also necessary, even for a level of intelligence I would call "direct vector intelligence," much below the level of a reptile, let alone a mouse.
To get so far as a reptile, I would claim the hardware also needs a kind of general and flexible but hard-wired MULTIPLEXING capability, to handle spatial complexity. But still implemented within the general capabilities that an NGNN can offer.
Beyond that, it's a couple of steps more to the basic mouse brain intelligence, but we will be very lucky if EITHER computational neuroscience or technological practice truly master these basic capabilities and requirements in our lifetime, at the general and fundamental level we need to claim some kind of full understanding. I've talked about what comes next, and put out a few equations on the relevant principles (yea even unto the possibility of building an intelligent computer based on quantum computing principles) , but it seems that it's important not to overload the information bit rate of the society we live in.
Best of luck,
Paul
-------- Original Message --------
Date: Mon, 12 Apr 2010 20:16:45 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Paul Werbos <pwerbos@nsf.gov>
CC: Minoru Asada <asada@ams.eng.osaka-u.ac.jp>, Angelo Cangelosi <acangelosi@plymouth.ac.uk>, Nik Kasabov <nkasabov@aut.ac.nz>, Giorgio Metta <pasa@dist.unige.it>, Asim Roy <ASIM.ROY@asu.edu>, Ron Sun <rsun@rpi.edu>, John Taylor <taymore2002@aol.co.uk>, Paolo Arena <parena@diees.unict.it>, Narayan Srinivasa <nsrinivasa@hrl.com>, Janusz Starzyk <starzykj@gmail.com>
Subject: Re: what neural networks can do
Paul and others:
Primarily as a recognizer for symbolic patterns or clean signals, the FSM and many symbolic systems such as HMM, Conditional Random Field (CRF) and Bayeisan Nets have bypassed the attention problem. Although artificial neural networks or connectionist cognitive models use emergent representations, they have been largely used as recognizers for a clean background for a lack of intent. Some networks (e.g., Weng et. al. Creception IJCV 1997, Fei-Fei Li PAMI 2006, Serre & Poggio PAMI 2007, Geoerge and Hawkins PLos Computational Biology 2009) have been used to detect a subset of patterns in a larger scene, but they still do not have an intent and an intent-directed reasoning capability to deal with real-world situations like in the 2-object+background example in my email.
Our developmental program and its specializations Where-What Networks deal with the above 2-object+background example. They indicate:
(1) A brain-like network can self-generate its intent (emergent).
(2) The intent is the origin of top-down attention.
(3) The top-down attention disregards backgrounds and distractors.
(4) The top-down attention is a primary mechanism for brain-like reasoning. For example, reasoning for location from a location intent and reasoning for type from a type intent.
(5) The time period where artificial neural network cannot reason well (as Marvin Minsky's AI Magazine paper criticized) is over.
It seems also reasonable to predict:
(a) Quantum phenomena in neuronal computation might be useful for simulating some details, but not necessary for the above 2-object+background example.
(b) The detailed shape of a neuron (Re: ours is a "point" neuronal model) might be useful for simulating detailed behavior of a neuron in fine time scales, but necessary for the above 2-object+background example.
(c) Spiking neuronal models might be useful for fine time scales, but not necessary for the above 2-object+background example.
During our panel session, I will present some detail about our Where-What Networks for the above 2-object+background example. Please let me know if any of you think that my above points have major flaws. Comments are also welcome.
I hope that the above 2-object+background example is useful for us to discuss the representation issue, between bottom-up and top-down.
-John
-------- Original Message --------
Date: Mon, 19 Apr 2010 14:37:57 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Yaochu.Jin@honda-ri.de
CC: suppressed
Subject: Please, please do not repeat the past mistakes that scientific history has made
--- Anonymous Review for the file 3 provided to some people on this email list as privileged material ---
In the introduction, the authors offends basically everybody and claims that no-one, in particular not reviewers, stand up to him and can understand anything: they are not competent. That is not science, at best it is polemics, at worst just ignorance.
The paper makes an rather endless number of exaggerated claims, which are not at all substantiated by the material presented. Some examples: claim: the SASE model "for the first time" deals with time in the correct way -- nothing to show that; claim the model deals "for the first time" with realistic vision -- only toy example shown, no quantitative results, all recent work and literature is not cited; claim the model performs better than Deep Belief -- nothing to substantiate ... and so on. Also: except self citations no really recent work is cited, which is in some contrast with the introduction claiming the the author has the world expertise in all relevant areas. Interestingly, he also claims that robotics is essential for a comprehensive theory, which he claims to develop in the paper, but the presented work has not connection to robotics at all, it ignores everything related to hardware ... There are many more issues could be discussed. So, based on scientific terms, I believe the paper must be rejected.
---------- End of review ---------------------
Dear Yaochu and all:
I will post this review on our discussion page: http://www.cse.msu.edu/ei/IJCNN10panel/ which has already heated email discussion.
Could you serve as the bind link between me and this respected reviewer so that he can keep his anonymity? Please forward my email and the above IJCNN10panel URL to him and ask him to join in our discussion through you. Please do the same for other two reviewers who have not submitted their reviews yet.
This discussion is very important for the large brain-mind community. I will take the flame and blame, but this process is useful for this large scientific community.
If you think that peer review works all the time, this is not what Thomas Kuhn documented based on his extensive study. Many senior experts said that peer review is conservative. In fact, it is more than conservative per Thomas Kuhn. I quote from http://en.wikipedia.org/wiki/Thomas_Kuhn to save your time:
----------- Quote from http://en.wikipedia.org/wiki/Thomas_Kuhn ---------------
Kuhn has made several important contributions to our understanding of the progress of knowledge:
* Science undergoes periodic "paradigm shifts" instead of progressing in a linear and continuous way
* These paradigm shifts open up new approaches to understanding that scientists would never have considered valid before
* Scientists can never divorce their subjective perspective from their work; thus, our comprehension of science can never rely on full "objectivity" - we must account for subjective perspectives as well
------------ End of quote from http://en.wikipedia.org/wiki/Thomas_Kuhn ---------------
I provide my reasoning and rebuttal to the material written by this respected reviewer. My material is not to him as a person. I would like to thank him to spend his time on this invited feature article.
> In the introduction, the authors offends basically everybody and claims that
> no-one, in particular not reviewers, stand up to him and can understand anything: they are not
> competent. That is not science, at best it is polemics, at worst just ignorance.
As Thomas Kuhn documented, this reviewer did not divorce his subjective perspective from his work. More, he linked the material even with his competency. No, I did not say that he is not competent.
What I wrote in the introduction is about a lack of infrastructure for confidently evaluate a grand picture of brain-mind. I trust you that this blind reviewer is probably among the best peers to review the picture we presented, but we predict that he will not feel "confident". Why?
In 1610 when Galileo Galilei published this account in favour of the sun-centered, Copernican theory, how many experts could confidently evaluate his account? The infrastructure was lacking! It is not the fault of anybody. It is simply that humans were not ready for this discovery.
In 1838 when Charles Darwin conceived his theory of natural selection, how many experts could confidently evaluate his theory? The infrastructure was lacking! Again, it is not the fault of anybody. It is simply that humans were not ready for this discovery.
> The paper makes an rather endless number of exaggerated claims, which
> are not at all substantiated by the material presented.
This reflected the situation I wrote about in the Introduction section: "fragmentation" of the brain-mind-intelligence field. The related major results have already published at different venues as cited in this invited feature article he reviewed. But he did not bother to read the papers that do substantiate. (You already stated that this manuscript is long.) The reviewer wanted such a grand picture with 5 chunks to be "substantiated" in a single feature article. Do you think that it is possible? He severely underestimated the space required to "substantiated" the 5 chunks. The main purpose of an invited feature article like ours is not to "substantiate". Instead, it is to summarize our latest published results to give a bigger picture as this society journal is designed for. I have provided CC to CIS officers.
> claim: the SASE model "for the first time"
> deals with time in the correct way -- nothing
> to show that;
The "first time" was used at four locations in the manuscript. Two cases are about time. I quote here exactly "(3) This is the first time that a connectionist model has dealt with time of any length without any dedicated temporal components. (4) This is the first time for a single connectionist model to deal with time warping and time duration, respectively, according to task context."
The reviewer's criticism "the correct way" is not accurate. We did not write "THE correct way". However, I stand by my exact text as I quoted above. I repeat: This is the first time, categorically in the above exact sense.
We have experimental results that showed that such a network dealt with long video sequence, long stereo sequences, and English sentences such as those from the Wall Street Journal. In the manuscript, we provided the text to guide the reviewer to find the "show":
"The SASE model has been tested for recognizing temporal visual events (Luciw et al. 2008 [55]), spatiotemporal disparity from stereo without explicit stereo images matching Solgi & Weng 2009 [78], text processing as temporal sequences for generalization to
new sentences based on synonyms, the part-of-speech tagging problem and the chunking problem using natural languages from the Wall Street Journal [94]."
Here are the links to facilitate this respected reviewer:
[55]: http://www.cse.msu.edu/~weng/research/ICDL08TemporalVision.pdf
[78]: http://www.cse.msu.edu/~weng/research/ICDL09TemporalStereo.pdf
[94]: http://www.cse.msu.edu/~weng/research/ICDL09TemporalText.pdf
For theoretical proof, see:
J. Weng, "A General Purpose Brain Model For Developmental Robots: The Spatial Brain for Any Temporal Lengths," in Proc. IEEE International Conference on Robotics and Automation, Workshop on Bio-Inspired Self-Organizing Robotic Systems, Anchorage, Alaska, May 3-8, 2010.
http://www.cse.msu.edu/~weng/research/ICRA10subm.pdf
In the manuscript, we have discussed positively the work of Dean Buonomano at UCLA (CCed). He is a neuroscience expert who argued for "intrinsic dynamics of nondedicated neural mechanisms" as Richard B. Ivry (CCed) at UC Berkeley reviewed in their recent review article in Trends in Cognitive Science 2008. Although Dean Buonomano did not use motor signal as a critical means for the brain to deal with both time duration issue (when the time duration is important and relevant to actions) and time warping issue (when the time duration between events is irrelevant to actions), the position of Dean Buonomano was correct, even though he seems a minority in neuroscience! Dean, your position is a "strange animal" in AI :) but I am with you as we chatted before.
> claim the model deals "for the first time" with realistic
> vision -- only toy example shown, no quantitative
> results,
Not an accurate criticism. I quote here exactly the 3rd "the first time": "(2) This is the first general purpose developmental model for recognizing general objects from complex backgrounds." I stand by this statement.
"Toy example"? Should we expect that the first work like that to show large data tests ...? This is another point I wrote in the Introduction section: Suddenly, the review criterion is not an advance of the-state-of-the-art, but the reviewer's wish-list.
"No quantitative results" is incorrect. Quantitative results are shown in Fig. 10(c), --- above 95% recognition rate (disjoint tests) with under 1.3 pixel average location error, for 5 realistic objects (cat, pig, truck, duck, car) with large viewing angle variation (disjoint tests) and 400 possible pixel locations. I repeat: This is the first time, categorically in the above exact sense.
> all recent work and literature is not cited;
This is not the first time in many reviews that I have received and I guess that it will not be the last time either. It gave a very bad "image" to your mind as a co-editor or the program manager of a government program, knowing that he would probably not have time to verify and confirm. Do you think that the reviewer should give at least a few examples to show that what he suggested are major misses for this line of work? Like many cases in the past, he gave none.
> claim the model performs better than Deep Belief -- nothing to substantiate ...
The Deep Belief network is good work and I cited under positive light. However, the Deep Belief network does not do attention and does not deal with complex backgrounds. "Nothing to substantiate"? Attention for backgrounds is "nothing"? I wonder whether he truly knows about vision. This is a holly grail! I respectfully proposed that finding relevant information from the "sea" of irrelevant information (visual background, background music and fans etc., the feel of your clothes, etc.) is a very much neglected essence of intelligence --- brain's and machine's alike.
> Also: except self citations no really recent work is cited,
Repetition, but worse: "no really recent work". See my rebuttal above.
> Interestingly, he also claims that robotics is essential for a comprehensive
> theory, which he claims to develop in the paper, but the presented work
> has not connection to robotics at all,
Narrowly considered, in this reviewer's view, robotics must involve a hardware robot. If he knows robotics well, he would have known that robotic hardware is not the most challenging issue for intelligent robots, but the brain-mind issue is. For example, the Location Motor in the Where-What Networks, WWN-1, WWN-2, WWN-3, is for driving a robot arm and robot drive base. The Type Motor in the WWNs is for driving the robot's "vocal tract" to tell what the object it sees.
I will be happy to add the above explanation into the manuscript.
Best regards,
-John
-------- Original Message --------
Date: Sun, 25 Apr 2010 14:15:36 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: suppressed
Subject: An example of lack of infrastructure
I respect this reviewer very much, as I guess that this reviewer is probably one of the most respected and well known modelers of brain-mind systems. I raise the points here just to indicate how badly the current infrastructure is serving our domain leaders who serve as reviewers to decide acceptance and funding. Humans are not ready for the arrived, exciting brain-mind age.
> Strong points of the paper:
> The studied question is very important, and the model itself is interesting.
Thank you. Our respected panels of federal agencies have stated similar comments. But, they wrote "but this proposal does not fit this panel" or something like that. Read further for some reasons.
> Major weak points, Observed deficiencies, and suggestions on how to improve them:
>
> I have several concerns regarding the style of the presentation. The paper
> starts with a nice, long introduction, but then unfortunately the author has no
> space left to explain many important details of his new model.
This is an conference paper for a special session on mental architecture and mental representation. It has a space limit of 8 pages. The section Introduction ends at the 7th line of page 2.
Section II. THE MENTAL ARCHITECTURES follows immediately at 8th line of page 2.
Section III. BRAIN SCALE: INTEGRATED PATHWAYS starts from page 3;
Section IV. PATHWAY SCALE: BOTTOM-UP AND TOP-DOWN from page 4;
Section V. CORTEX SCALE: PRESCREENING from page 4;
Section VI. LAYER SCALE: THE DUALLY OPTIMAL LCA from page 5;
Section VII. REASONING WITH THE CONCRETE AND THE ABSTRACT from page 7;
Section VIII. EXPERIMENTS from page 7;
Section IX. CONCLUSIONS AND DISCUSSION from page 8.
This paper presents a picture of brain-mind puzzle after our twenty years of research with many technical papers published elsewhere. This is the first time when a grand picture is cautiously presented. But our respected reviewers expect that they should be able to fully understand such a picture from a single paper or a single proposal (say, 15 pages). This respected reviewer called much of the work simply "long introduction". Dear reviewer, mathematical detail is useful here, but it is only a small part of the picture. The brain-mind issue is way beyond pure mathematics can handle.
I respectfully raised that a confident evaluation of such a picture of brain-mind requires breadth-and-width in biology, neuroscience, psychology, electrical engineering, computer science and mathematics. I hope that the planned Brain Mind Institute as a service institution can get all your support.
> In the whole section the mathematical details should be more precise. ...
> We do not know how z(t_n) and x(t_n)
> evolves over time. Also, the paper is not self-contained in the sense that LCA
> has not been explained; only a reference was given for it.
The reference was: J. Weng and M. Luciw, "Dually Optimal Neuronal Layers: Lobe Component Analysis," IEEE Transactions on Autonomous Mental Development, vol. 1, no. 1, pp. 68-85, 2009.
http://www.cse.msu.edu/%7Eweng/research/LCA-IEEE.pdf
Our respected reviewers consistently do not think that following (at least browsing) the cited references is their responsibility as a reviewer.
This is a systemic problem: Lack of guidelines and bylaws of checks-and-balances for in the current peer review process.
During 1787, our founding fathers of the US already knew clearly how important checks-and-balances is for the Autonomous Mental Development (AMD) of this great country called US. But as this great country has prospered through our founding father's wisdom, our later generations became less sensitive to their vision. Peer review, the developmental program of our science and technology, has not had sufficient checks-and-balances.
The problem is not peer review itself. The current peer review is not just conservative, as Alan Leshner and some respected NSF Program Manager have stated. The systemic problem of peer review is the surprising lack of bylaws for the entire process of peer review! Adding text for two specific questions in the proposal summary is useful but is superficial and insufficient. The current lack of bylaws for the peer review process is like lack of developmental program for this country's science and technology.
Humbly yours,
-John
-------- Original Message --------
Date: Mon, 26 Apr 2010 13:01:19 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: suppressed
Subject: Why all 6 fields are necessary for confidently understanding the true grand picture of brain-mind?
Dear All:
I proposed the six fields that are necessary for confidently understanding the brain-mind picture:
Biology, Neuroscience, Psychology, Electrical Engineering, Computer Science and Mathematics.
The follow is my brief account based on my personal experience:
(i) Biology: how the cell works tells us how the brain wires itself through experience. For example, I got a guidance about how the brain works from the genomic equivalence principle. Without knowing it, your agent model is likely wrong and computationally ineffective. As another example, neuromodulation will tell us why our emotion and reinforcement learning studies are still shallow and they are just partial story for the brain. It also helps us to understand why at the top level, the 5 chunks of the brain-mind picture are more basic than neuromodulation and many other detail.
(ii) Neuroscience: Many models from psychology and machine learning and many psychological experimental designs can be improved if the proposers knew about neuroscience. Personally, I have found that the brain ways are more computational effective than almost all the psychological and machine algorithms that I am aware of. The latter is typically more complex but less computationally effective. For example, the original proposers of HMM, Bayesian nets, support vector machines, and many neural network models would have dropped their models if they had known how the brain does what he wants.
(iii) Psychology: We need psychology to know many facts about how the animals and humans learn and behave, why the brain does not have many "inborn" behaviors, why some psychologists still believe that a human baby has inborn knowledge about the physical world (they are wrong, but their experiments and arguments are useful for us to think about the brain's developmental program), and why children's language learning requires so few examples. The fact that our brain-mind model can learn from just a few samples in high dimensional parameter space and many other properties are guided by such rich psychological evidence.
(iv) Electrical engineering. It tells many important concepts that the brain deals with for its large practical problems when it deals with high-dimensional signals, images, video, muscles and glands. Personally, although my degree is from computer science, my advisers Narendra Ahuja and Thomas S. Huang are professors in electrical and computer engineering at UIUC. My background in signals and systems has a lot to do with my MS and PhD years. Why is it so difficult for an AI researcher in computer science to understand how the brain works? Why can they not go beyond symbols? I have seen many such cases. We need to change.
(v) Computer Science. The automata and complexity theory in computer science enables one to understand why many existing neural network methods do not do well. AI methods, although traditional, tell us what is goal-directed reasoning and what are required for the brain to do goal-directed reasoning. Many neural network researchers have a
background in electrical engineering. A background in computer science can help to understand that recognition and detection are not sufficient for the brain and why.
(vi) Mathematics. Believe or not, the brain is a highly optimal processor in order to do all the tasks in life using a such a limited-size brain. Without the knowledge about some major results and tools in probability and statistics, we at MSU would not have been able to come up with the Lobe Component Analysis (LCA) for modeling brain's feature development, which out-performed some well known, neural-inspired, popular techniques. For understanding how the brain processes information, one needs mathematics to see the deeper reasons.
Just my 2 cents of worth.
Best,
-John
-------- Original Message --------
From: Professor Ron Sun <rsun@rpi.edu>
To: Juyang Weng <weng@cse.msu.edu>
Cc: suppressed
Subject: Re: Why all 6 fields are necessary for confidently understanding the true grand picture of brain-mind?
Date: Tue, 27 Apr 2010 12:35:36 -0400
hi, John:
interesting list. however, I think you are missing the disciplines concerning sociocultural aspects, which are indispensable to brain-mind development.
regarding sociocultural aspects of cognition, please see:
* R. Sun, L. A. Coward, and M. J. Zenzen, On levels of cognitive modeling . Philosophical Psychology, Vol.18, No.5, pp.613-637. 2005. [Formatted PDF]
see also:
* R. Sun, Cognition and Multi-Agent Interaction: From Cognitive Mdoeling to Social Simulation. Cambridge University Press, New York. 2006.
best,
-- ron
-------- Original Message --------
Date: Tue, 27 Apr 2010 20:11:50 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Professor Ron Sun <rsun@rpi.edu>
CC: suppressed
Subject: We are mislead by the apparent complexity of the brain-mind
Hi Ron:
Thank you very much for raising this sociocultural aspect of modeling. Minoru Asada also raised this point. He even added body to model. I was not very direct to Minoru, due to part of my weakness of human nature. I was thinking about that you and Minoru are among the leaders who have done great jobs. Furthermore, you two are probably my closest friends and allies. Let us just talk about our common interests, and down play our differences. I know that this "down play" is not good for research.
Let us all first agree on our common interests. We are allies.
But we face the grand challenge of brain-mind. Many people have already lost confidence of clearly understanding the brain-mind issue!
To make our discussion productive, please allow me to be direct this time. My apology for being direct. This way, everybody on this list can understand the key points without taking too much their time. I believe their time is well spent to read this email.
- Marvin Minsky wrote in the prologue of his book "Society of Mind": "Perhaps the fault is actually mine, for failing to find a tidy base of neatly ordered principles. But I'm inclined to lay the blame upon the nature of the mind: much of its power seems to stem from just the messy ways its agents cross-connect. If so, that complication can't be helped; it's only what we must expect from evolution's countless tricks."
- David Marr proposed three levels of modeling: computation, algorithms, and implementation.
- You proposed four levels of modeling: substrates, intra-agent processes, agents, inter-agent processes. Apparently you added inter-agent processes as another level of modeling.
- Minoru Asada proposed body should also be part of modeling. According to your idea, then, Minoru should also model body differences and multi-body configurations among different people, baby, toddler, child, mid-age, elder, handicapped and so on.
- CYC has proposed further: ontology. All human common sense knowledge should be modeled by handcrafted models.
This track is overwhelming and has been done many times in AI and Cognitive Science. But it has proved to be brittle in real world. Government agencies have lost confidence but some disguised programs of this track are still leak into our federal programs. More disappointments from this track will follow, I predict.
Why? Many deep reasons. The most obvious one is the run-away exponential complexity that has to be handled by human hands. I plan to prove it in a future paper.
I have spent some time on biology, developmental psychology, neuroscience, and developmental neuroscience. I realized that we are mislead by the apparent complexity of the brain-mind.
Why? The first chunk of my picture of the brain-mind directly addresses this misled: the genome (developmental program) of our brain is "task nonspecific". This means that it applies to any task, any single person, multiple persons, various human bodies, various social settings and various countries. All four levels you want to hand-model are emergent, automatic! This is the miracle of development.
Many people heard about our Workshop on Development and Learning 2000. They started to read development in psychology and neuroscience. Overwhelming details! They are overwhelmed and gave up.
Our picture of brain-mind is about the principled science that Marvin Minsky liked to get.
In a recent email communication that I had with David Mumford, I wrote about my "very simple model" "It is only about a building block of the brain-mind. Interestingly, such building blocks autonomously build the brain's sky scrapper. Few can believe that it is how the brain does to build the mind."
This simple model has the 5 chunks integrated: development, architecture, area, space and time.
The brain-mind picture has more details, such as neuromodulation (which takes care of reinforcement learning and emotion), but they are not as fundamental as the above 5 chunks. Yes, it takes care of your 4 levels autonomously,without your handcrafting. I will present this very simple model during the IJCNN 2010 Special Session on Mental Architecture and Representation. During our panel session, I can only briefly outline it.
By the way, one of my colleagues in neuro-psychology commented: "It is simple to draw a diagram" but he retracted politely.
Yes, it looks simple, but it took me and my many students twenty years. I think that it is what Marvin Minsky liked to get.
Humbly yours,
-John
-------- Original Message --------
From: Professor Ron Sun <rsun@rpi.edu>
To: Juyang Weng <weng@cse.msu.edu>
Subject: Re: We are mislead by the apparent complexity of the brain-mind
Date: Tue, 27 Apr 2010 20:58:35 -0400
Cc: suppresed
On Apr 27, 2010, at 8:11 PM, Juyang Weng wrote:
> Every four levels you want to hand-model is emergent, automatic!
> This is the miracle of development.
who said anything about "hand-model"? (what is it anyway?)
many cognitive models are capable of learning -- even autonomous learning and development (to some extent). Your model was not the first one.
What I would like to see is exactly the kind of model capable of autonomous learning but with sufficiently broad functionalities.
> This simple model has the 5 chunks integrated: development,
> architecture, area, space and time.
> The brain-mind picture has more details, such as neuromodulation
> (which takes care of reinforcement learning and emotion), but they
> are not as fundamental as the above 5 chunks. Yes, it takes care
> of your 4 levels autonomously,without your handcrafting.
at this point, any model can "autonomously" take care of all 4 levels (from neurobiology to social processes)? this must be "miracle"! it is definitely news to me. are you planning on proving it or at least demonstrate it with plenty of empirical evidence?
-------- Original Message --------
Date: Tue, 27 Apr 2010 21:35:32 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Professor Ron Sun <rsun@rpi.edu>
CC: suppressed
Subject: Re: We are mislead by the apparent complexity of the brain-mind
Your question indicates a wide-spread phenomenon, but please be patient. Let me explain.
A handcrafted representation typically have learning (including reinforcement learning), such as Minky's work, CYC, yours and Minoru Asada's work. For example, consider HMM and Bayesian nets, which are two general-purpose learning frameworks. But they use handcrafted representations. Pick HMM as example: A human designer picks up n phonemes and m words. At the phoneme level, n HMMs are trained, one for each phoneme. At the word
level, m HMMs are trained, one for each word. Then, the world level takes inputs from the phoneme level. What are the handcrafted representations? The internal (inside the skull) "walls" among n phonemes, the internal
"walls" among m words, and the internal "walls" between the phonemes set and the word set.
The brain does not do that, as the brain does not "care" your human designed concepts such as phonemes and words. The brain exists first, before humans have symbolic concepts such as "phoneme" and "word". In all our published models, internal features are all emergent. In fact, none of the features in our model (inside the skull) can be described by any human linguistic terms precisely.
-John
-------- Original Message --------
Date: Tue, 27 Apr 2010 22:12:04 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: Juyang Weng <weng@cse.msu.edu>
CC: suppressed
Subject: Re: We are mislead by the apparent complexity of the brain-mind
> any model can "autonomously" take care of all 4 levels (from neurobiology to social processes)?
> this must be "miracle"! it is definitely news to me.
>
are you planning on proving it or at least demonstrate it with plenty of empirical evidence?
This is basically the biology is doing for every higher animal, such as vertebrates, mammals, primates and humans. From a developmental program in the single cell to a fully functional human brain-mind (and body by
the way), that displays the 4 levels of phenotype you want to handcraft. Yes, all autonomously.
Of course, conventional artificial neural networks also use fully emergent representations. However, they have great problems, including the ones that Marvin Minsky criticized since 1980s.
Our publications at different venues have shown, collectively, that the 5 chunks of the brain-mind have been identified, which solve the problems that Marvin Minsky criticized as special cases. Of course, with the limited resource that we have in the MSU EI lab, those displayed at the 4 levels are still simple. But the proof of feasibility has been done with our theory and our experimental results.
However, as soon as I cautiously put all these accepted results together as a picture of the brain-mind, my respected reviewers got upset. One reason is probably that they do not follow our publications, like you, understandably. This is what I meant by the scientific history repeating itself.
Best regards,
-John
-------- Original Message --------
Date: Wed, 28 Apr 2010 09:00:53 -0400
X-Google-Sender-Auth: 7efc55f1ea3d28cf
Message-ID: <v2qd5729c391004280600l62498a82vd6a664d0d9c7c4f7@mail.gmail.com>
Subject: Re: Why all 6 fields are necessary for confidently understanding the true grand picture of brain-mind?
From: Paul Werbos <werbos@ieee.org>
To: suppressed
Ron's comment touches on something very important.
In my view, the most important realistic challenge to real science -- mathematically well-grounded consensus exploration -- in this century is to truly understand the higher levels of intelligence which exist in the brain of the individual mouse. i.e. its ability to learn to "predict" its environment and to perform better and better in the decision challenges it faces.
It is ALSO important that humanity continue to try to understand ourselves, and levels of the mind beyond the mouse brain level, but that is a different kind of enterprise(s), not so ready for unified mathematics or consensus except in a few islands. It is not a good idea to try to impose consensus where it does not exist.
Issues like collective intelligence, the foundations of language, Jungian synchronicity, mirror neurons and "Gaia" certainly have an important bearing on humanity's effort to understand itself and to understand the mind in general. But to do maximum justice BOTH to up-to-mouse-brain science AND to beyond-mouse-brain analysis, it may be important to keep the distinction clear. The two areas can and should inform each other, but that may actually be easier if we keep the distinction in mind.
By the way, I have heard that "collective intelligence" has become a topic of extreme interest across Eastern Asia lately, at all levels. EQ chapter two?
Best of luck,
Paul
P.S. This is not an official view of NSF, of course. Also, I am not replying to the ENTIRE list; please forgive me for
a relatively random cutoff. Presumably there should either be a listserv created, which people can opt out of
easily, or we should not continue many more rounds of this.
-------- Original Message --------
Date: |
Sun, 25 Jul 2010 14:59:15 +0200 |
From: |
Juyang Weng <weng@cse.msu.edu> |
To: |
Angelo Cangelosi <acangelosi@plymouth.ac.uk>, Nik Kasabov <nkasabov@aut.ac.nz>, Giorgio Metta <pasa@dist.unige.it>, Asim Roy <ASIM.ROY@asu.edu>, Ron Sun <rsun@rpi.edu>, John Taylor <taymore2002@aol.co.uk>, "Juyang (John) Weng" <weng@cse.msu.edu>, Paul Werbos <pwerbos@nsf.gov>, Paolo Arena <parena@diees.unict.it>, Narayan Srinivasa <nsrinivasa@hrl.com>, Janusz Starzyk <starzykj@gmail.com>, Edgar Koerner <Edgar.Koerner@honda-ri.de>, Yan Meng <yan.meng@stevens.edu> |
Subject: |
Presentation slides have been posted |
Dear All:
Thank you all for the contribution to our panel.
Your presentation slides have been posted on our panel web site:
http://www.cse.msu.edu/ei/IJCNN10panel/
You can also directly download from the URL:
http://www.cse.msu.edu/ei/IJCNN10panel/2010-7-21-IJCNN-Rep-Panel.pps
Please feel free to continue our discussion on these important issues
via emails.
Best regards,
-John
-------- Original Message --------
Date: Mon, 13 Sep 2010 19:44:03 -0400
From: Juyang Weng <weng@cse.msu.edu>
To: supressed
Subject: A shorter version of the brain-mind picture
Dear Domain Experts:
As I expected, the version I sent to you for your input was rejected (too long was one of the reasons). This is a shorter version that was published in IJCNN 2010: 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: http://www.cse.msu.edu/~weng/research/WCCI10rvsd.pdf
Here is a nutshell from this model:
The model predicts that the brain --- including the spinal cord, the hindbrain, the midbrain, and the forebrain --- is “patched” through evolution by many general purpose brain areas. The working of each area can be understood as a “bridge” representation that assists its two “banks”. This model has been supported by a series of experimental studies with several functions that did not have clear prior computational solutions, such as connectionist abstraction, concept emergence, goal-directed top-down attention, and connectionist reasoning. This brain-mind picture requires five (5) necessary "chunks": development (how the brain-mind emerges), architecture (how areas connect), area (how each building block represents), space (how the brain deals with spatial information) and time (how the brain deals with temporal information). This brain-mind model puts two large schools into the same picture --- the symbolic representation school and the connectionist school. The experimental studies include visual attention, visual recognition, stereo (without explicit binocular matching), and early language learning/generalization.
Some of you have contacted me individually along this line. Your inputs are very much appreciated.
Best regards,
-John
Top | Abstract | Organizers | Qestions | Tentative Panelists | Email Discussions
To Weng's Home Page: http://www.cse.msu.edu/~weng/