Deep physical neural networks trained with backpropagation

Abstract

Deep-learning models have become pervasive tools in science and engineering. However, their energy requirements now increasingly limit their scalability1. Deep-learning accelerators2–9 aim to perform deep learning energy-efficiently, usually targeting the inference phase and often by exploiting physical substrates beyond conventional electronics. Approaches so far10–22 have been unable to apply the backpropagation algorithm to train unconventional novel hardware in situ. The advantages of backpropagation have made it the de facto training method for large-scale neural networks, so this deficiency constitutes a major impediment. Here we introduce a hybrid in situ–in silico algorithm, called physics-aware training, that applies backpropagation to train controllable physical systems. Just as deep learning realizes computations with deep neural networks made from layers of mathematical functions, our approach allows us to train deep physical neural networks made from layers of controllable physical systems, even when the physical layers lack any mathematical isomorphism to conventional artificial neural network layers. To demonstrate the universality of our approach, we train diverse physical neural networks based on optics, mechanics and electronics to experimentally perform audio and image classification tasks. Physics-aware training combines the scalability of backpropagation with the automatic mitigation of imperfections and noise achievable with in situ algorithms. Physical neural networks have the potential to perform machine learning faster and more energy-efficiently than conventional electronic processors and, more broadly, can endow physical systems with automatically designed physical functionalities, for example, for robotics23–26, materials27–29 and smart sensors30–32.

Device-system Co-design of Photonic Neuromorphic Processor using Reinforcement Learning

“…We demonstrate a few examples of utilizing the developed optical accelerator in the applications of image classification and materials discovery. The prediction accuracy in these applications is improved through physics-aware training process 28 . Furthermore, we show that the developed O-GEMM hardware accelerator can be further employed in the RL algorithms to accelerate the chip design of another optical accelerator.…”

“…Although the material and device parameter space in our demonstration is not gigantic, the developed highly-parallelized hardware emulator enables the further exploration of various RL algorithms in large-scale optimization problems. Moreover, the demonstrated physics-aware training approaches lay out strategies on how to deploy physical hardware systems to different ML application scenarios 28 .…”

Learning by non-interfering feedback chemical signaling in physical networks

“…Indeed, there have been a myriad of scholarly attempts to construct learning algorithms within purely physical systems [21][22][23][24][25][26]. Here, we will focus on "equilibrium propagation" [24,27] and "coupled learning" [25].…”

Physics-aware Complex-valued Adversarial Machine Learning in Reconfigurable Diffractive All-optical Neural Network

“…The inevitable loss of phase information during these conversions leads to a substantial measurement deviation from the computer-developed model. Additional adjustment and iterative training for error corrections are needed 15,22 , and the reconfigurability and transfer to other tasks are still laborious to accomplish. Multiple conversions between electrical and optical domains also weaken the advantages of employing optical approaches over electrical ones.…”