Competitive Optical Learning with Winner-Take-All Modulators

Modern neural network learning models such as competitive learning networks[2,4], resonance correlation networks, and back propagation networks[1] require a wider range of neuron behavior than a simple saturating threshold non-linearity. However, optical implementation of neurons that incorporate non-local, non-linear functions such as shunting inhibition, winner-take-all, and history-dependent behavior is beyond the capability of conventional optical devices. A new class of light modulator has been developed that combines the flexibility of analog and digital electronic VLSI circuits, optical detectors, and the switchable electo-optic capabilities of liquid crystal materials. In this paper we will show how these liquid crystal/VLSI modulators can be used in optical implementations of these learning networks. We discuss in detail a competitive optical learning network which uses LC/VLSI winner-take-all neurons on fractal grids to program adaptive volume holographic interconnections. We will present results from tests of the LC/VLSI winner-take-all modulator arrays, and in addition will show preliminary results from an optical competitive learning system that uses the LC/VLSI modulators as neurons.