100520 Schrader

The Bernstein Center Freiburg

Bernstein Seminar

Sven Schrader

HONDA Research Institute Europe GmbH
Offenbach, Germany

A columnar model of bottom-up and top-down processing in the neocortex

Abstract:
In vision, an object representation is constructed through a cascade of areas, each processing more and more elaborate features of the stimulus image. We hypothesize that the necessary signal flow involves a feed-forward path to compose the object representation as well as a feedback path to refine the representation and to correct errors. Using word recognition as a metaphor for visual stimuli, we have implemented this architecture in a spiking neural network. Given a string of characters, the network quickly generates an initial stimulus hypothesis after only 20-30 ms. Already the first wave of action potentials traveling upwards the hierarchy contains the correct local symbols (syllables, word) of the presented stimulus primitives. In most cases, the activation of the correct word precedes all other candidates with millisecond precision and is most pronounced. Thus, our model codes the correctness of a response in its relative latency. In a subsequent refinement stage, where high-level activity modulates lower stages, this activation dominance is propagated back, influencing its own afferent activity to exhibit a unique decision.

Host: Arvind Kumar