Cognition

High Performance Computing for Learning

The most effective approaches to problems in learning are computationally intensive, using optimization techniques in high- dimensional spaces with hundreds or thousands of parameters and simulation of neurobiological and developmental systems, and in addition require a grasp of realistic and cognitive constraints. Integrating machine perception and learning into the HPCC environment provides an opportunity to develop more robust and flexible practical systems for tasks such as visual inspection and can aid understanding of how the brain works.

Under some conditions, image examples can be used directly (without physically-based three-dimensional models) for image analysis and graphics synthesis using learning techniques. With this approach, many example images of an object under different imaging conditions (such as poses and expressions) are gathered, and an "analysis network" learns to associate to every image a vector of parameters such as pose and expression. A "synthesis network" learns the inverse task of producing the corresponding image given certain pose and expression parameters. When only one example image is available, new "virtual" images can be generated by learning from prototypical examples, as illustrated in the image below. This learning exploits a class of multidimensional interpolation networks that approximate the nonlinear mapping between vector input and vector output. In addition to face recognition, application areas include computer graphics, special effects, very low bandwidth teleconferencing, interactive multimedia, and object recognition systems.

Conducted at the Center for Biological and Computational Learning at MIT, this work is funded by NSF, other Federal organizations, and corporations.

http://www.ai.mit.edu/projects/cbcl/HPCCBlueBook/BlueBook.html

The central image is the original camera shot and the surrounding images were generated from the original using image synthesis/analysis.

A New View of Cognition

Understanding how the brain directs natural behavior requires changing the environment on millisecond timescales at key points in the on-going task. To study this problem, simulated worlds have been generated by special purpose high performance computing systems interfaced with head-mounted visual displays and visual and kinetic monitoring equipment.

This virtual reality approach funded by NIH has already led to unexpected results. Recent findings concern the brain's ability to rapidly switch its focus and absorb just the right amount of information for its current task. This process is called temporal fractionation of physical properties.

Observations of the common properties of physical objects show that they cohere; for example, the color and location of an object are unified in normal human perception when timescales are equal to or longer than a second. Behavioral experiments using innovative virtual reality equipment have found that the brain processes color and relative location at two different times during the course of solving the task. This result was entirely unanticipated, since the color and location of an object are unified in normal human perception. This new view of cognition revises theories of how the brain prepares information for problem solving.

Other virtual reality research includes hand-eye virtual environments and driving simulators. These projects are yielding interesting research results and have promising applications in addressing problems such as perception and distance deficiencies in elderly drivers. These studies will also serve as diagnostic aids for a wide variety of illnesses such as Huntington's and Parkinson's diseases that are characterized by short-term memory deficits.