Large-scale neuromorphic optoelectronic computing with a reconfigurable diffractive processing unit
This work proposes the reconfigurable diffractive processing unit, an optoelectronic fused computing architecture based on the diffraction of light, which can support different neural networks and achieve a high model complexity with millions of neurons.
Abstract
Application-specific optical processors have been considered disruptive\ntechnologies for modern computing that can fundamentally accelerate the\ndevelopment of artificial intelligence (AI) by offering substantially improved\ncomputing performance. Recent advancements in optical neural network\narchitectures for neural information processing have been applied to perform\nvarious machine learning tasks. However, the existing architectures have\nlimited complexity and performance; and each of them requires its own dedicated\ndesign that cannot be reconfigured to switch between different neural network\nmodels for different applications after deployment. Here, we propose an\noptoelectronic reconfigurable computing paradigm by constructing a diffractive\nprocessing unit (DPU) that can efficiently support different neural networks\nand achieve a high model complexity with millions of neurons. It allocates\nalmost all of its computational operations optically and achieves extremely\nhigh speed of data modulation and large-scale network parameter updating by\ndynamically programming optical modulators and photodetectors. We demonstrated\nthe reconfiguration of the DPU to implement various diffractive feedforward and\nrecurrent neural networks and developed a novel adaptive training approach to\ncircumvent the system imperfections. We applied the trained networks for\nhigh-speed classifying of handwritten digit images and human action videos over\nbenchmark datasets, and the experimental results revealed a comparable\nclassification accuracy to the electronic computing approaches. Furthermore,\nour prototype system built with off-the-shelf optoelectronic components\nsurpasses the performance of state-of-the-art graphics processing units (GPUs)\nby several times on computing speed and more than an order of magnitude on\nsystem energy efficiency.\n