Experimental results are reported for an optical associative memory previously described. This system is a single-layer neural network architecture simulating a 2-D array of approximately 105 neurons on which images can be represented. This array is fully interconnected by holograms, and the system is organized as an autoassociative memory with feedback. An external image projected into the system causes one of the stored images to become a stable state of the system. The ability of the system to recognize distorted versions (e.g. rotated, shifted, or scaled) of a stored image depends critically on the gain of the system as the light goes around the loop. High gain provides invariance to distortions but ultimately it also leads to a loss in discrimination against unfamiliar images. Thus there is an optimum choice of parameters of the system that yields optimum performance. A description is given of how the parameters affect the performance of the memory, and the performance (in terms of discrimination vs. invariance) obtained by the experimental system is reported.
|Number of pages||8|
|State||Published - 1 Dec 1988|