Research & Projects

The foundations of our work have their roots in the early papers on the ontogenesis of neural circuits by Professor von der Malsburg. This work outlines a theory of how connections in the brain develop. The problem of neural ontogenesis plays a central role in models of the brain. There are approximately 10^15 (a million billion) synaptic variables in the brain. How can we possibly hope to adjust a set of parameter values of this order of magnitude in simulations of brain function? On the other hand, if the set of rules or algorithms that determine how brains develop were known, these could be used to get brain models to set their own parameters in a self-organizing fashion.

Work on neural ontogenesis leads naturally into a powerful class of neural networks known as the "dynamic link architecture". The dynamic link architecture tries to extract causal relations in sensory input. It does this using correlations in neural firing patterns and rapidly modifiable synaptic weights which respond to these correlations. In one way, the dynamic link architecture may be thought of as an ontogenetic solution to the problems of information processing in mature neural circuits. We are studying the dynamic link architecture as a possible model of brain function.

Finally, we use an algorithmic version of the dynamic link architecture called "elastic graph matching" to solve various problems in computer vision. This group is currently focusing on problems associated with automatic perception and recognition of human faces. See also the article that appeared in the USC Chronicle.