All-optical nonlinear activation function based on stimulated Brillouin scattering

TL;DR

This paper demonstrates an all-optical nonlinear activation function using stimulated Brillouin scattering, enabling deeper, more efficient photonic neural networks without opto-electronic conversion losses.

Contribution

It introduces a coherent, frequency-selective, tunable in-fiber activation function based on stimulated Brillouin scattering, enhancing photonic neural network capabilities.

Findings

Achieved an optoacoustic activation function tunable between LeakyReLU, Sigmoid, and Quadratic.

Provided signal amplification up to 20 dB to offset insertion losses.

Demonstrated potential for deep optical neural networks.

Abstract

Photonic neural networks have demonstrated their potential over the past decades, but have not yet reached the full extent of their capabilities. One reason for this lies in an essential component - the nonlinear activation function, which ensures that the neural network can perform the required arbitrary nonlinear transformation. The desired all-optical nonlinear activation function is difficult to realize, and as a result, most of the reported photonic neural networks rely on opto-electronic activation functions. Usually, the sacrifices made are the unique advantages of photonics, such as resource-efficient coherent and frequency-multiplexed information encoding. In addition, opto-electronic activation functions normally limit the photonic neural network depth by adding insertion losses. Here, we experimentally demonstrate an in-fiber photonic nonlinear activation function based on stimulated Brillouin scattering. Our design is coherent and frequency selective, making it suitable for multi-frequency neural networks. The optoacoustic activation function can be tuned continuously and all-optically between a variety of activation functions such as LeakyReLU, Sigmoid, and Quadratic. In addition, our design amplifies the input signal with gain as high as $20\,\mathrm{dB}$, compensating for insertion losses on the fly, and thus paving the way for deep optical neural networks.