Photorefractive crystals as trainable neural network interconnects
Michael J. O'Callaghan, Dana Z. Anderson · Annual Meeting Optical Society of America · 1988
We have performed experimental studies of a hybrid electronic/optical neural network where interconnections are implemented optically using a photorefractive crystal. This system1 consists of a laser, liquid crystal television (LCTV) screen, photorefractive crystal, and CCD light detector array to implement a two-layer feed-forward neural network. A microcomputer reads the CCD array and controls the LCTV. The pixels of the LCTV represent a set of input neurons (which spatially modulate a laser beam). A holographic grating within the crystal forms the neural network interconnections (via diffraction) and the CCD array corresponds to a set of output neurons. The holographic grating is trained using adaline learning, also known as the delta rule.2 A 1-D training image (64 pixels wide) modulates a laser beam which then passes through the crystal. A computer compares the spatial profile of diffracted light seen at the 256-element CCD array with the desired output. A 64-pixel wide corrective image modulates a second laser beam which intersects the first, and the grating evolves in response to the overlapping images. This procedure is repeated many times for each input image to be learned.