Application of Reconfigurable All-Optical Activation Unit based on Optical Injection into Bistable Fabry-Pérot Laser in Multilayer Perceptron Neural Networks

TL;DR

The paper addresses the need for nonlinear, reconfigurable activation units in all-optical neural networks while ensuring cascadability. It proposes an activation unit based on injection-locked bistable Fabry-Pérot lasers, whose nonlinear transfer function $P_{\text{out}} = Φ(P_{\text{in}})$ is tunable via detuning $\Delta\omega$ between the injected signal and a side mode. Offline TensorFlow training is used to determine layer weights for a two-hidden-layer perceptron, and numerical simulations of optical signal propagation validate the approach on MNIST and Fashion-MNIST, yielding up to about $95\%$ and $85\%$ test accuracies, respectively. The work demonstrates that reconfigurable all-optical activation units can enable high-performance, low-power photonic neural networks without additional inter-layer amplification, with activation shapes adapted to the task at hand.

Abstract

In this paper we theoretically investigate application of a bistable Fabry-Pérot semiconductor laser under optical-injection as all-optical activation unit for multilayer perceptron optical neural networks. The proposed device is programmed to provide reconfigurable sigmoid-like activation functions with adjustable thresholds and saturation points and benchmarked on machine learning image recognition problems. Due to the reconfigurability of the activation unit, the accuracy can be increased by up to 2% simply by adjusting the control parameter of the activation unit to suit the specific problem. For a simple two-layer perceptron neural network, we achieve inference accuracies of up to 95% and 85%, for the MNIST and Fashion-MNIST datasets, respectively.

Figures (4)

• Figure 1: a) Schematic of reconfigurable activation unit architecture based on FP-LD under optical-injection. b) Free-running spectrum of FP-LD under optical injection and in the case of injection-locking. c) Optical peak power transfer function of FP-LD for different values of $\Delta\omega$. d) Two-hidden-layer perceptron architecture.
• Figure 2: Averaged testing a) accuracy and b) loss for MNIST and d) accuracy for Fashion-MNIST classification for all combinations of available activations in the hidden ($\Delta\omega_1$) and in the output layer ($\Delta\omega_2$). c) The same as a) for two-hidden-layer perceptron with 25 units in the hidden layer.
• Figure 3: The evolution of the accuracy (left axis, blue lines) and losses (right axis, red lines) during the training of the model for activation pairs corresponding to high ((-29$\Omega$, -29$\Omega$), solid lines) and low (-38$\Omega$, -15$\Omega$), dashed lines) training accuracy.
• Figure 4: Activations used for training (solid lines) and peak power values (blue dots) of actual optical signals propagating through (a) hidden and (b) output layer for ($-29\Omega, -29\Omega$). Insets: examples of optical signals waveforms at the input (dashed lines) and output (solid lines) of activation units.