A continuum model for a neural network with lateral inhibition biased near threshold behaves like a dispersive optical medium with convective gain in a narrow band of wave numbers. A quadratic memory nonlinearity causes a long-lived change in an excitability parameter with continued excitation, imprinting multiplexed volume holographic gratings in the medium. Pairs, Chains, and Cycles of associated images may then be reconstructed by multiple Bragg scattering, corresponding to the recognition of patterns of images invariant under groups of transformations. An optical analog of the neural medium is suggested which could implement pattern recognition tasks on noncoherent inputs within a single crystal.
Marcus Cohen, Marcus Cohen,
"Distributed Computation In Neural Networks And Their Optical Analogs", Proc. SPIE 0540, Southwest Conf on Optics '85, (20 November 1985); doi: 10.1117/12.976169; https://doi.org/10.1117/12.976169