Photonic Implementations of Neural Networks

Abstract

Several broad classes of neural networks comprise distributed, nonlinear, dynamical systems in which large numbers of relatively simple processing elements (neuron units) are densely interconnected. The interconnections are often configured such that the interconnection weights are adaptive and contain the learned memories and behaviors of the system. Advanced optical interconnection techniques are being developed that can potentially be used in conjunction with optoelectronic neuron units to implement photonic neural-like computational modules ( e.g ., Fig. 1) with relatively large array sizes (10 5 to 10 6 neuron units) and a high degree of connectivity (fan-outs and fan-ins of 10 4 to 10 6 , with 10 9 to 10 12 total interconnections). A key open question is whether the high bandwidths (potentially 100 MHz or more) available from hybrid optoelectronic spatial light modulators (SLMs) can be effectively combined with such high density volume holographic optical interconnections (dynamically recorded in photorefractive materials) to provide enhanced computational throughput capacity as well as complex neural network simulation capability. A second key open question is whether advanced electronic/photonic packaging technologies can provide capability for system-level integration of highly compact multichip modules that exhibit both local (multi-plane) and global interconnections (Fig. 2).