The continuing growth of unstructured digital media content in the form of video, audio, and sensor data is driving a need for more effective and efficient methods for indexing, searching, categorizing and organizing such information. Additional complexity to this problem is created by the need to provide results as fast as possible. The final challenge stems from the fact that the computing resources are always limited and finite. In this proposal, we seek to develop novel machine learning algorithms that enable real-time video and sensor data analysis on large data streams given limited computational resources. We focus on healthcare as our application domain where real-time video analysis can prevent user-errors in operating medical devices or provide immediate alerts to caregivers about dangerous situations.
A fundamental challenge in developing real-time video and sensor analysis for applications such as healthcare arises from the mismatch between the firehose of data, the high complexity of the prediction algorithms used for complex analysis and the limited computational resources that are available on the device itself, at a patientsí home or even at a medical facility. To address this research challenge, we propose to develop machine learning theories and algorithms that automatically adapt to hardware, with the aim to learn, from a large number of training examples, a prediction function that (i) is sufficiently accurate in making effective predictions and (ii) can be run efficiently on a specified computer system to deliver time critical results. Three types of prediction models are studied in the proposal for the problem of automatic hardware adaptation, including the vector-based model, the matrix-based model, and the prediction model based on a function from a Reproducing Kernel Hilbert Space (RKHS). We propose a general framework and multiple optimization techniques to learn accurate prediction models that match limited memory and computational capacity. The developed learning algorithms will be evaluated in several medical scenarios through real-time prediction of a patient's activities from observations in the large video archives collected by the CMU CareMedia and related projects.