Neuromorphic photonics with electro-absorption modulators.

Jonathan K George,Paul R Prucnal,Armin Mehrabian,Alexander N Tait,Jiawei Meng,Thomas Ferreira De Lima,Bhavin J Shastri,Rubab Amin,Volker J Sorger,Tarek El-Ghazawi

doi:10.1364/oe.27.005181

Jonathan K George, Paul R Prucnal + Show 8 more

Open Access

https://doi.org/10.1364/oe.27.005181

Copy DOI

Abstract

Photonic neural networks benefit from both the high-channel capacity and the wave nature of light acting as an effective weighting mechanism through linear optics. Incorporating a nonlinear activation function by using active integrated photonic components allows neural networks with multiple layers to be built monolithically, eliminating the need for energy and latency costs due to external conversion. Interferometer-based modulators, while popular in communications, have been shown to require more area than absorption-based modulators, resulting in a reduced neural network density. Here, we develop a model for absorption modulators in an electro-optic fully connected neural network, including noise, and compare the network's performance with the activation functions produced intrinsically by five types of absorption modulators. Our results show the quantum well absorption modulator-based electro-optic neuron has the best performance allowing for 96% prediction accuracy with 1.7×10-12 J/MAC excluding laser power when performing MNIST classification in a 2 hidden layer feed-forward photonic neural network.

Highlights

Photonic neural networks (NN) have the potential for both high channel capacity and low operating power. The former is provided via a) charing of small electrical capacitors and b) ’bosonification’ where many photons are allowed to occupy the same quantum state, such as technologically utilized in wavelength division multiplexing (WDM)
In our analysis of cascaded SNR we assume that the input power to each neuron during operation is uniformly distributed in the nonlinear portion of the voltage transfer function
The MNIST dataset [13] is a set of images of handwritten digits in a grayscale 28x28 pixel format that is commonly used in comparing neural network performance

Summary

Introduction

Photonic neural networks (NN) have the potential for both high channel capacity (i.e. data baud rate) and low operating power. In ring-based weighting, photonic rings are selectively tuned to apply a dot product, a potentially more compact method with wavelength-division multiplexing (WDM) In both cases linear optics achieves an efficient weighting where the wave nature of light computes the inner product by propagating forward in time. In the electro-optic neuron [8] considered here consisting of a photodiode connected either directly, or through an interface circuit, to an optical modulator, the nonlinear response of the modulator itself can be used to generate the necessary nonlinearity in the signal transfer function In this case, the choice of modulator type will immediately impact the shape of the activation function and the operation of the network. Our results show the quantum well absorption modulator based electro-optic neuron outperforming the other three absorption modulators across a wide range of modulator lengths and optical powers

Modulators for Nonlinear Activation

Photodiode Coupling

Noise and Cascadability

Results and Discussion