Abstract
(1) Introduction. The use of diffractive optical elements (DOE’s) and microlens arrays in optical interconnection systems can potentially provide high throughputs and small system volumes, using components that are amenable to automated design and mass production techniques. This paper considers fixed-weight neural network interconnections based on such components, and focuses on the realm of small system volumes and short propagation lengths (~1 mm) with potential for cascading into a compact, multilayer free-space system. We consider the space-variant interconnection system of Fig. 1, and achieve short propagation lengths by restricting each fanout pattern to a local neighborhood. The system uses an array of NxN sub-DOE's at the input plane to connect to an array of NxN detectors at the output plane. For this locally connected neural network interconnection, each sub-DOE stores one weighted fanout pattern that connects to MxM nearest neighbors in the output plane. The beam incident on each sub-DOE comes from a modulator or an emitter (not shown), which represents an interconnection input node (e.g., an output of a neuron unit). The optics of the interconnection system provides a Fourier transform (in magnitude) from each sub-DOE to the detector array, which serves as a set of neuron unit inputs. We constrain the sub-DOE spacing to be equal to the detector spacing to allow for use in multilayer systems. A globally connected space-variant system can be realized similarly by replacing the microlens array with a bulk lens and letting each input node connect to all output plane detectors [1,2]. The minimum propagation distance and system volume can be shown to be approximately proportional to M and N^M for the local system, and N and A3 for the global system, respectively. While the local system can be used in a smaller volume, its crosstalk levels can be high due to reconstruction noise (e.g., diffraction orders outside of the local fanout neighborhood) of each sub-DOE. In this paper, we describe a crosstalk reduction method enabled by varying the mapping from reconstructed spot locations to detector locations. Several novel DOE designs for, and simulations of, local fixed-weight neural network interconnections are evaluated as a test of this method.
© 1995 Optical Society of America
PDF ArticleMore Like This
Takayuki Ishida and Masatoshi Ishikawa
PD1 Optical Computing (IP) 1995
Ahmed Louri, Hongki Sung, Yoonkeon Moon, and Bernard P. Zeigler
PThC5 Photonics in Switching (PS) 1995
W. Lee Hendrick, Philippe J. Marchand, Frederick B. McCormick, Ilkan Çokgör, and Sadik C. Esener
OThB4 Optical Computing (IP) 1995