A novel model of two-dimensional image associative memory for optical neurochips

A novel type of optical neurochip, which associates complete stored images in response to the incomplete inputs, is presented The original point of this work is that the optical neurochip allows direct image processing in terms of parallel input/output interface and parallel neural processing. So the processing speed of 100 times higher than the conventional neurochips is obtained. The operation principle is based on the nonlinear transformation of the input two-dimensional images to the corresponding one-dimensional vectors. A Hebbian-like learning rule is utilized in order to make the whole system simple. The simulation results verify that the system has very high error-collection and ultra-fast processing capabilities.