Analysis and Evaluation of Electro-Optic Artificial Neural Network Performance in the Presence of Non-Ideal Components

Artificial neural networks (ANNs) are usually designed around vector-matrix multipliers, where the inputs to the neurons are represented by the vectors while the interconnection weights are represented by the matrix. Optics, with its interferenceless free-space communication capabilities, is therefore an efficient and natural way to implement ANNs; however, it is not without practical problems. In electro-optic ANNs, errors due to non-linear or limited accuracy components could exist in the input devices, the interconnection weight matrix, or the output detectors. This report addresses some of those errors in terms of a specific implementation and across several ANN architectures. The electro-optic neuralnetwork used for this research consists of two electro-optic ANN layers and one ANN layer implemented in software only. In the electro-optic layers, light emitting diodes (LEDs) are used to provide the input, liquid crystal spatial light modulators (SLM) serve as the interconnection weight matrixes, and photodiode detectors are the nonlinear thresholding elements. Specific hardware imperfections—nonuniform LED illumination, optical misalignment and cross talk within the SLM, thermal drift in the SLM, and noise and linearity problems in the photodiode detectors circuits—are analyzed and experimentally documented. The impact of these errors on the performance of an ANN is dependent upon the ANN architecture. These error sources, as they effect the design of this or any other electro-optic ANN, are then discussed and evaluated for several representative ANN architectures. Many decisions must be made when designing a practical implementation of an electro-optic ANN. This work provides some basis of how such decisions may be made.