|
Numerous applications in science and technology nowadays utilize deep learning to tackle challenging computational tasks. With the increasing demand for deep learning, high-speed and energy-efficient accelerators are urgently needed. Although electronic accelerators are flexible, optical computers holds great promise due to their potential for massive parallelism and low power consumption. However, optical computing platforms demonstrated so far have mostly been limited to relatively small-scale computing tasks, despite the potential for scalability. Here, we propose and demonstrate a hardware-efficient design that allows deployment of a reconfigurable deep neural network (DNN) architecture without a direct isomorphism to standard DNN designs. Our proposed system is scalable and supports larger-scale computing. Our system realizes an optical neural network (ONN) using a digital micromirror device (DMD) for encoding data and trainable parameters, a complex medium for random complex weight mixing, and a camera for nonlinear activation and optical readout. A straight-through estimator enables backpropagation, even with a DMD as a binary encoding device. With this ONN as an elementary building block and automating the search for neural architectures, we can build complex and deep ONNs for a range of large-scale computing tasks, such as 3D medical image classification. The architecture-optimized deep ONNs are deployed by time-multiplexing data streams in one system. Our system enables large-scale training and inference in situ. Furthermore, we demonstrate that our system is capable of achieving task accuracies close to that of state-of-the-art benchmarks with more complex architectures implemented in silico.
|
