Investigation of charge coupled circuits as an enabling technology for photonic neural networks

The research presented in this dissertation represents an experimental investigation into silicon device architectures and circuits which may be utilized to implement powerful electro-optical neural networks. The architectures investigated combine the parallel input of optical signals with efficient implementation of electrical signal processing to create an imaging transversal filter which can efficiently compute outer-products, a fundamental processing kernel for neural network processing. The focus of this work is the investigation of, specifically, charge coupled circuits as an enabling technology for photonic neural networks. In the context of neural network processing, the devices developed acquire two dimensional matrix coefficients by integrating incident light patterns. Simultaneously, these devices introduce binary or ternary weighting patterns through parallel electrical inputs. Upon completion of the integration of the spatial light patterns and the introduction of the electrical weights the product of the incident light and the electrical weight is formed at each pixel location. This operation is followed by a rapid sum-of-products computation along one dimension of the two dimensional array. In this fashion an outer product is formed. The sum-of-products is computed in the charge domain and serially read-out to a single analog output port through a charge coupled device (CCD). Two generations of the transversal imager are presented. The designs incorporate buried channel CCDs, surface channel CCDs, charge primed couplers, charge switching shift registers, floating gate modulating circuits, floating diffusion amplifiers, photocapacitors, and NMOS dynamic shift registers. Detailed analysis and simulations performed on the various circuits utilized in the two generations of transversal imager devices are presented. Comparisons of these circuit simulations with actual test device measurements are presented. The performance of these charge coupled circuits as photonic neural networks is evaluated for the Hopfield content addressable memory (CAM) paradigm. The transversal imager technology is also compared with other neural network enabling technologies. A comparison, with respect to neural network processing power, is made with other CCD neural networks, analog VLSI networks, and digital VLSI neural networks. This research has verified that the transversal imager concept is sound and demonstration of a 32 x 32 transversal imager with ternary (+1, 0 $-$1) weighting is presented. It was shown that the transversal imager may be operated in two modes, (1) optical input of the binary network state vector and electrical input of the ternary synaptic weights, or (2) electrical input of the state vector and optical input of true analog synaptic weights. This very powerful capability in a single programmable device is thought to be unique, and is a significant achievement of this research