Scalable reservoir computing on coherent linear photonic processor

Abstract

Photonic neuromorphic computing is of particular interest due to its significant potential for ultrahigh computing speed and energy efficiency. The advantage of photonic computing hardware lies in its ultrawide bandwidth and parallel processing utilizing inherent parallelism. Here, we demonstrate a scalable on-chip photonic implementation of a simplified recurrent neural network, called a reservoir computer, using an integrated coherent linear photonic processor. In contrast to previous approaches, both the input and recurrent weights are encoded in the spatiotemporal domain by photonic linear processing, which enables scalable and ultrafast computing beyond the input electrical bandwidth. As the device can process multiple wavelength inputs over the telecom C-band simultaneously, we can use ultrawide optical bandwidth (~5 terahertz) as a computational resource. Experiments for the standard benchmarks showed good performance for chaotic time-series forecasting and image classification. The device is considered to be able to perform 21.12 tera multiplication–accumulation operations per second (MAC ∙ s−1) for each wavelength and can reach petascale computation speed on a single photonic chip by using wavelength division multiplexing. Our results are challenging for conventional Turing–von Neumann machines, and they confirm the great potential of photonic neuromorphic processing towards peta-scale neuromorphic super-computing on a photonic chip.

Reservoir computing based on a silicon microring and time multiplexing for binary and analog operations

“…The overall power consumption is estimated to be around , and is distributed as follows: for the Pump and Probe lasers, for the electrical amplifiers, for the AWG, for the optical amplifiers, for the oscilloscope and for the laptop. This is comparable to the value of reported in a recent work on a silica chip 29 , and is expected to not significantly differ from the one of more advanced photonic classifiers 8 , 9 , 43 , since most of the power consumption arises from the electronics used for data acquisition, control and post processing, rather than from the optical components. The architecture is proof of concept and suffers from an unusually slow carrier recombination time.…”

“…It is natural to seek the use of hybrid spatio-temporal architectures as an optimal trade-off solution. These may adopt a small scale spatial topology with sparse connectivity, where each physical node can accommodate few hundreds of time multiplexed virtual nodes, similar to what is suggested in 28 and recently demonstrated on a silica chip 29 . It is worth noting that there are other approaches to increase the dimensionality of the network.…”

Wave-based extreme deep learning based on non-linear time-Floquet entanglement

“…These classification accuracies are competitive. For example, they are higher than the ones reported in reference 43 (parallel image classification). Also, the classification-accuracy results are comparable with other relevant works despite the decreasing the pixel sizes of all images 46 .…”

“…This memory effect allows RNNs to detect recursive relations in the data, which is relevant for example to process temporal signals. In digital implementations, however, the heavy internal connectivity matrices that are involved in the training process make RNNs particularly computationally-expensive and complicated [43][44][45][46] . In order to solve these challenges, a number of alternative computing approaches such as long short-term memory (LSTM) 47 , echo state networks (ESNs) 48 , extreme learning machines (ELMs) [49][50][51] , and reservoir computing (RC) [44][45][46]52 have emerged.…”

Simulating graphene dynamics in synthetic space with photonic rings

“…The modulated ring systems can be realized in either fiber-based system [41][42][43][44] or on-chip lithium niobate resonator 45 , which brings our proposal to a flexible experimental setup with state-of-art technologies in bulk optics or integrated photonics. Our work not only broadens the current research on synthetic dimensions in photonics 26,27 , but also enriches quantum simulations with topological photonics 46,47 , which shows potential applications in optical signal processing 48,49 and quantum computations 50,51 .…”