End-to-end optical training of neural networks

University of Oxford / Physics

Machine learning, one of the most revolutionary emerging technologies, has dramatically reshaped the landscape in different fields and completely transformed the way we approach complex tasks.

Artificial neural networks (ANNs) form the foundation of modern machine learning techniques. The training of ANNs, however, requires a massive amount of computation that is both time and energy consuming, and quickly saturates the available computational power. This poses a severe challenge to the electronic computation architecture, especially with the declining of Moore’s law. A possible solution is offered by optics, which is a natural platform for linear computing thanks to its coherence and superposition properties.

Dr Xianxin Guo

Although optical neural networks (ONNs) have been investigated for a long time, optical training of neural networks has never been realized. The training of neural networks involves (1) signal forward propagation;

(2) loss function calculation;

(3) error backpropagation and

(4) weight parameter update.

While the first two steps can be implemented with optics in different ways, the backpropagation step (3) is extremely challenging since it requires a) different responses of the nonlinear unit in forward and backward directions and b) derivative calculation of the nonlinear function. This optical backpropagation problem has been hindering the development of ONNs for decades.

Recently, we discovered that optical backpropagation can be realized by adopting the simple pump-probe scheme with saturable absorbers. With forward-propagating signal and backward-propagating error taking the roles of strong pump and weak probe beams, all the requirements of backpropagation are satisfied. Although optical backpropagation is an outstanding challenge to this field, our solution is surprisingly simple and effective, and it allows us to build ONNs with end-to-end optical training for the first time. As we found through numerical simulation in image classification tasks, our ONN can reach state-of-the-art performance with practical experimental parameters.

In this proposed research project, I aim to realize the first optically-trained ONN with the optical backpropagation scheme. The ONN will be constructed with free-space optics and atomic vapor cells, and it will be benchmarked in image classification tasks. The long-term goal is to build ONNs that can outperform current electronic ANNs.