Neuron unit arrays and Nature/Nurture adaptation for photonic multichip modules

To implement a previously proposed 3-D hybrid electronic/photonic multichip module (PMCM) (mimicking a primate retina structure) capable of low-latency, high-throughput, parallel-processing computations, several critical hardware components are designed, fabricated, and tested. All components are made of MOSIS 1.5 mum n-well BiCMOS (bipolar complimentary metal oxide silicon) fabrication process.; A 12-by-12 dual-input, dual-output silicon neuron unit array chip has been fabricated, and characterized. A desired sigmoid-shape optical output from a vertical surface emitting laser (VCSEL) driven by this chip (with a linear-optical-input) was obtained. A logarithmic amplifier circuitry has been fabricated, and characterized. The dynamic range of its sensed brightness is multiple decades wide. This bipolar-based circuit's high sensitivity at low input signal range can improve the overall optical responsivity of the PMCM if it is integrated. A floating gate design is verified to be a good candidate for the long-term analog weight storage. The floating gate controlled channel resistance can represent the lateral weighted interconnection in the PMCM. A preliminary active pixel sensor design is also characterized, and evaluated for weight storage. Physical constraints, trade-offs, and relationships among the components for optimizing the performance of the PMCM are discussed.; Software-wise, an artificial neural learning algorithm (Nature/Nurture algorithm) is developed for modeling the PMCM. This algorithm describes the weight updating rules for both the vertical fixed (nature-like) and the lateral adaptive (nurture-like) weighted interconnections in the PMCM. The learning algorithm for the lateral weight adaptations is new, and derived based on the multi-layer error back-propagation (BP) supervised learning algorithm using gradient descent method. Results from a simple optical character recognition (OCR) simulation show: (1) A PMCM with only one hidden neuron layer is sufficient to perform the OCR. (2) The Nature/Nurture trained neural network can recognize well the new modified patterns (generated from the original patterns) after the lateral weight adaptations. (3) A neural network similar to the pathways of the PMCM with local connectivity (only 9 vertical and 8 lateral interconnections from each neuron) can also perform pattern recognition with acceptable recognition rate.