SELF-ORGANIZING NEURAL NETWORK MODEL OF THE VISUAL CORTEX

Joseph Sirosh

---

Scientific Significance:

The goal of this project is a unified self-organizing neural network model of the visual cortex, in which both visual feature detectors and connections that associate visual features develop from visual input. Several novel results have been obtained to date, resulting in two technical reports which have been submitted for publication, one for a prominent conference and the other to a journal. The most significant discovery is that Ocular Dominance Columns and Lateral Connectivity in the primary visual cortex can develop synergetically and simultaneously by a very simple Hebbian self-organizing process and that the same simple process can account for much of the plasticity (such as dynamic receptive fields) observed in the mature visual cortex. Thus, a unified explanation has been put forth for early cortical development and mature cortical plasticity.

Current research is extending the frontiers even further. Large scale simulations are being done to show that (1) the same self-organizing mechanism can account for orientation columns as well, (2) lateral connections learn associations between feature detectors such as simple Gestalt rules and (3) several visual illusions arise as a result of such self-organization. If successful, the research will give a highly unified and concise description of cortical development and function.

---

Numerical Approach and Performance:

The approach uses standard self-organizing neural networks. The simulation algorithm is highly parallel, but requires fast communication and fast synchronization. The T3D has proved to be exceptionally useful for this work because of its shared memory and hardware synchronization features.

The scaling is indistinguishable from linear.

30 MFLOPS/PE
32 PE = 8.5 Equivalent C90 CPU's

---

Back to Contents Page

---