BIOGRAPHY
What are the basic principles of representation and computation in the nervous system? Cognitive scientists have argued for a theory based upon compositionality, which refers to the evident ability of brains to represent objects, scenes, thoughts and actions in a hierarchical structure. I am studying a mathematical formulation for compositionality, and the implications of this formulation for interpreting neural activity patterns and for building computer vision systems. If the theory is correct, then the neural representation of a perceived object such as the letter "E" would include representations of the object's parts -- three horizontal strokes and one vertical stroke. A basic tenet of the theory is that the activity pattern among the hundreds of millions of neurons that participate in the representation of the whole can be teased apart into the activity patterns of the parts. The representation of the parts would be more or less independent of the particular whole - these are reusable parts. The part/whole relationship is hierarchical, and would presumably describe the gluing together of local features (edges, surface patches, and so on) recursively into descriptions of entire scenes. The theory predicts a statistical "signature" in the activity patterns of neurons signaling a perceived part/whole relationship. We are attempting to confirm this prediction. The theory also suggests a framework for building an artificial vision system, and we are attempting to build such a system.
|
