Only-train-electrical-to-optical-conversion (OTEOC): simple diffractive neural networks with optical readout

Abstract

Machine learning hardware based on optical diffraction is emerging as a new computing platform with high throughput and low latency. The current all-optical diffractive deep neural networks often suffer from complex optical configuration, lack of efficient optical nonlinear activation, and critical alignment between optical layers for system integration. The opto-electronic diffractive neural networks can partially address these issues by shifting some computation load, e.g., nonlinear activation and adaptive training, to the electronic domain. However, these hybrid networks require extra optical-to-electrical conversion that inevitably slows the overall process down. Here, we propose a simple opto-electronic diffractive neural network with just one optical layer enabled by a standard phase-only spatial light modulator. The proposed system can classify images by optical readout and does not need to collect the light distribution for subsequent electronic computation. The nonlinear function is intrinsically integrated in the essential encoding process from the electronic input to the modulated wavefront of light. Thanks to its simplicity, the system can reach high classification accuracy without calibration and can be reconfigured by updating the weights without changing or moving any physical component. We believe this technology brings diffractive neural networks a step closer to building realistic optics-based neurocomputers.