Single-Shot Optical Neural Network

As deep neural networks (DNNs) grow to solve increasingly complex problems, they are becoming limited by the latency and power consumption of existing digital processors. 'Weight-stationary' analog optical and electronic hardware has been proposed to reduce the compute resources required by DNNs by eliminating expensive weight updates; however, with scalability limited to an input vector length $K$ of hundreds of elements. Here, we present a scalable, single-shot-per-layer weight-stationary optical processor that leverages the advantages of free-space optics for passive optical copying and large-scale distribution of an input vector and integrated optoelectronics for static, reconfigurable weighting and the nonlinearity. We propose an optimized near-term CMOS-compatible system with $K = 1,000$ and beyond, and we calculate its theoretical total latency ($\sim$10 ns), energy consumption ($\sim$10 fJ/MAC) and throughput ($\sim$petaMAC/s) per layer. We also experimentally test DNN classification accuracy with single-shot analog optical encoding, copying and weighting of the MNIST handwritten digit dataset in a proof-of-concept system, achieving 94.7% (similar to the ground truth accuracy of 96.3%) without retraining on the hardware or data preprocessing. Lastly, we determine the upper bound on throughput of our system ($\sim$0.9 exaMAC/s), set by the maximum optical bandwidth before significant loss of accuracy. This joint use of wide spectral and spatial bandwidths enables highly efficient computing for next-generation DNNs.