The construction of a programmable multilayer analogue neural network using space invariant interconnects
N. Collings, Ali R. Pourzand, R. Völkel · Optical Computing · 1995
Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.
Optical Computing · Optica Publishing Group · conference-paper · 1995
Abstract
The use of a multilayer neural network is indicated in those cases of pattern classification where the input has a relatively low spatial complexity, eg 16 x 16 pixels. Such an input size arises in the post-segmentation stage of handwritten character recognition, or more generally after a pre-processing stage on more complex input scenes. Since the pre-processing is likely to be optical, eg Fourier or wavelet transform, it is of interest to consider the construction of an optical neural network where the training might be slow, due to the speed of the interface of the programmable weight matrices, but the classification stage would proceed at rates superior to electronics.
This involves the use of stand-alone analog optical device for the intermediate layer of neural thresholding elements (hidden layer) in between the two layers of interconnects. The critical aspects of such an approach are the engineering of the programmable interconnect, the characteristics of the hidden layer optical device, the question of optical subtraction, and the use of discretization techniques to avoid the deleterious consequences of analog noise. The first aspect will be discussed in this summary and the other aspects will be more fully reported at the conference.
The optical design of the system was presented previously [1], and this report will concentrate on the practical results.
Open the resolved locator · 10.1364/optcomp.1995.omc19
Metadata licence: OpenAlex metadata (CC0)