Large-scale neuromorphic optoelectronic computing with a reconfigurable diffractive processing unit

Application-specific optical processors have been considered disruptive technologies for modern computing that can fundamentally accelerate the development of artificial intelligence (AI) by offering substantially improved performance. Recent advancements in neural network architectures information processing applied to perform various machine learning tasks. However, existing limited complexity and performance; each them requires its own dedicated design cannot be reconfigured switch between different models applications after deployment. Here, we propose an optoelectronic reconfigurable paradigm constructing a diffractive unit (DPU) efficiently support networks achieve high model with millions neurons. It allocates almost all computational operations optically achieves extremely speed data modulation large-scale parameter updating dynamically programming modulators photodetectors. We demonstrated reconfiguration DPU implement feedforward recurrent developed novel adaptive training approach circumvent system imperfections. trained high-speed classifying handwritten digit images human action videos over benchmark datasets, experimental results revealed comparable classification accuracy electronic approaches. Furthermore, our prototype built off-the-shelf components surpasses performance state-of-the-art graphics units (GPUs) several times on more than order magnitude energy efficiency.