Training Deep Neural Networks in Situ with Neuromorphic Photonics

Photonic analog signal processing and neuromorphic computing [Invited]

“…Further works in photonic neuromorphic computing have been focused on solving this problem of implementing wider and deeper neural networks in hardware [15,18,19,22] . While output technology, output signal domain, and input/output signal behavior are all of some consequence to the performance of a PNN, by far the most promising area for improvement is the method of implementing weighted addition.…”

“…Their analog nature allows for signal processing in continuous time, and reduces the cost, memory requirements, and precision loss resulting from the need to digitize massive amounts of data. Not only are these advantages for applications involving optical signals, such as optical fiber or free space optical communication [2][3][4][5][6][7][8][9][10][11][12][13] , but they also show promise when applied to computation in general [14][15][16][17][18][19][20][21][22][23][24] .…”

“…All of these capabilities show potential for improving high-speed communication systems, as they require throughputs that would overwhelm most analog electrical circuitry, and drive up the price of digital signal processing techniques through increasingly high-speed analog-to-digital converters. They also show promise for accelerating general computation, as many of these devices are capable of being integrated into silicon wafers and used as application-specific co-processors in tandem with digital and analog electrical circuitry [18] . Photonic blind-source separation is a technique that uses photonic circuits to manage interference in wireless communication systems [2][3][4][5] , free-space optical communication systems [4] , and optics-based sensor technologies [29,30] .…”

“…Photonic neuromorphic computing is a field that aims to leverage the advantages of analog photonic circuits for highperformance computing tasks. In particular, it leverages the high speed, wavelength-division parallelism, and low power consumption of photonic arithmetic operations to accelerate the calculations that make up artificial neural networks [14,[16][17][18][19][20][21][22][23][24]27,28,[34][35][36][37] . This enables simple yet powerful networks to solve complex tasks at the speed light transmission [13] .…”

“…[14], where it used photonic spike processing to perform the operations of a spiking neural network. Since then, multiple techniques have been made to improve upon this design, including tunable weight banks made from interferometers [14,35] , tunable resonators [15][16][17] , or phasechange elements [22][23][24]34] , as well as a variety of output designs including all-optical thresholders [22,23] , electro-optical nonlinear laser gates [28,36] , and electrical sampling and reproduction circuitry [18,19,24] . Today, it is showing particular promise in the fields of high speed neural signal processing [13,37] and machine vision acceleration [19][20][21][22][23][24] .…”