Silicon photonic architecture for training deep neural networks with direct feedback alignment

Abstract

There has been growing interest in using photonic processors for performing neural network inference operations; however, these networks are currently trained using standard digital electronics. Here, we propose on-chip training of neural networks enabled by a CMOS-compatible silicon photonic architecture to harness the potential for massively parallel, efficient, and fast data operations. Our scheme employs the direct feedback alignment training algorithm, which trains neural networks using error feedback rather than error backpropagation, and can operate at speeds of trillions of multiply–accumulate (MAC) operations per second while consuming less than one picojoule per MAC operation. The photonic architecture exploits parallelized matrix–vector multiplications using arrays of microring resonators for processing multi-channel analog signals along single waveguide buses to calculate the gradient vector for each neural network layer in situ. We also experimentally demonstrate training deep neural networks with the MNIST dataset using on-chip MAC operation results. Our approach for efficient, ultra-fast neural network training showcases photonics as a promising platform for executing artificial intelligence applications.

Experimentally realized in situ backpropagation for deep learning in photonic neural networks

“…Population-based methods ( 42 ), direct feedback alignment ( 43 , 44 ), and perturbative approaches ( 16 ) have some advantages but are ultimately less efficient for training neural networks compared to backpropagation, especially for hybrid PNNs. Unlike “receiverless” fully analog PNNs ( 16 ), hybrid PNNs require optoelectronic (i.e., digital-analog and analog-digital) conversions for each layer, which can slow down perturbative training.…”

“…Such schemes may be compatible with mixedsignal schemes for accelerators that already aim to reduce the current communication energy bottleneck (39,41) in the race to address the energy-doubling AI problem (3). Population-based methods (42), direct feedback alignment (43,44), and perturbative approaches ( 16) have some advantages but are ultimately less efficient for training neural networks compared to backpropagation, especially for hybrid PNNs. Unlike "receiverless" fully analog PNNs (16), hybrid PNNs require optoelectronic (i.e., digital-analog and analog-digital) conversions for each layer, which can slow down perturbative training.…”

An Electro-Photonic System for Accelerating Deep Neural Networks

“…This degradation can be quite drastic even for DNN inference, especially for higher h. Furthermore, DNN training is often more sensitive to quantization noise than inference, as gradient calculation requires a relatively higher dynamic range. Therefore, the scope of DNN acceleration with photonic hardware has been limited to DNN inference except for a few works [5], [19], [24], [36], [58] that focused on training very small DNN models and simple tasks. Generally speaking, training state-of-the-art DNN models using photonic compute cores has been a far-fetched goal due to the limited precision of analog operations.…”

Reconfigurable focusing and defocusing meta lens using Sb2Se3 phase change material

“…Compare to other proposed on-chip integrated photonic hardware accelerator, in memory computational photonic tensor core 23,27,31,[34][35][36][37][38][39][40][41][42][43][44] , all optical convolution solution provides alternative solution with higher data processing rate and lower power consumption 45,46 . With the combine implementation of high performance electrooptic modulators [47][48][49][50][51][52][53][54][55][56] , photodetectors [57][58][59] , signal converters 60 and the reconfigurable metalens, a fully integrated all optical neuromorphic architecture could be realized soon. In this work, as proof as the concept, a reconfigurable lens based on a low loss phase change material Sb2Se3 61,62 optical properties transformation (∆n) between amorphous and crystalline states which are introduced by the ITO based novel low loss transparent heater is proposed and demonstrated.…”