Abstract

Massive interconnection is one of the key features of artificial neural networks, which can be implemented optically by encoding the interconnection weight matrix (IWM) on a spatial light modulator (SLM) and superimposing all the IWM-submatrix with a lenslet array. The low dynamic range and very limited resolution of the currently available SLMs are the main concern for the development of optical neural network (ONN). So far the largest ONNs with lenslet array contain only 64 neurons and 4096 interconnections,^1,2 which can have little practical use. However, the weakness can be effectively strengthened by a novel electron trapping (ET) material that has a resolution of ~100 lines/mm, dynamic range of 4 orders of magnitude, and is suitable for a 3-D interconnection implementation. An ONN architecture with 128 × 128 neurons is proposed in this paper. The massive interconnections are stored in 100 layers of ET thin film. The page-addressable 3-D ET memory offers a huge information capacity and may have vast uses in real-time target tracking and pattern recognition. The proposed system can store a total of 2.7 × 10¹⁰ interconnections and have a processing speed as high as 5.4 × 10¹¹ interconnections per second. System design considerations as well as performance evaluations are provided.

© 1991 Optical Society of America