Abstract
Based on our earlier work on the encoding of complex stimuli by cortical networks (Wyss et al., 2003) we designed a neural network to classify visual stimuli. The transmission delays of the connections among the excitatory neurons, which are proportional to length, result in a complex spatial and temporal activity pattern when a visual stimulus is presented to the network. The pooled neuronal activity, or temporal population code - TPC, represents a stimulus-specific fingerprint that is invariant to translations, rotations, and small distortions. The TPC thus allows a highly accurate and robust stimulus classification in situations where a single target is presented to the network. However, classification fails when the scene also includes a closely positioned distractor. We show here that by introducing dendritic attenuation in the model, we are able to increase classification performance in the presence of a distractor, i.e. segmentation. Moreover, by combining codes of several attenuation levels, we obtain a segmentation model that is robust and scalable both with and without distractor. Hence, our results suggest that both dense lateral connectivity and dendritic structure provide a complementary computational substrate for the encoding of complex stimuli.
This project is supported by the Swiss Federal Institute of Technology (ETH) Zurich.