Utilizing nonlinear fiber optics in neural network architecture

Abstract

Many types of machine learning algorithms have been developed, but neural networks have proven to be the most capable ones so far. For example, GPT models used in ChatGPT are highly complex and neural network based. So, neural networks are powerful tools in creating artificial intelligence and in solving other problems requiring machine learning. However, training neural networks takes vast amounts of time and electricity. To combat this problem different types of physical systems have been studied to replace the in silico calculations. These neural networks which use physical systems as a part of the network are called physical neural networks (PNNs). Simplest type of and the one used in this thesis is called physical reservoir computer (PRC). PRC contains a physical system, which the data is put through, and a trainable output layer.

In this thesis we looked this problem from the perspective of nonlinear fiber optics. There are a lot of different types of nonlinear phenomena, but in this thesis, we discuss three of them: self-phase modulation, supercontinuum generation and modulation instability. In short self-phase modulation widens the pulse spectrum depending on the input pulse parameters. Supercontinuum generation is combination of multiple different nonlinear effects including but not limited to self-phase modulation, anomalous dispersion, soliton fission and modulation instability. Modulation instability on the other hand is the pulses instability to hold its shape when perturbed.

The data used in training is MNIST data. It is 28 x 28 grayscale images of handwritten numbers from 0 to 9. There are 60 000 training images and 10 000 test images. These images were down scaled to 14 x 14 and reshaped to a vector of length 196. These vectors were then encoded to either the input pulses intensity or phase. The trainable output layer was trained to make the classifications based on the pulses after the propagation. The classification accuracy of the input pulses was also tested to determine if any improvement happened.

The resulting classification accuracies of PRCs with self-phase modulation, supercontinuum generation and modulation instability were 84%, 84% and 95% respectively. Input pulses with data encoded on the temporal intensity have classification accuracy of 92% and with data encoded on the temporal or spectral phase the input pulse classification accuracy is 84%. The only improvement was seen with the modulation instability. The classification accuracy of the modulation instability based PRC input pulses being so high makes it hard to determine if the system does anything remarkable. In future different datasets and PNN models could be used to determine if any advantage could be attained with PNNs with modulation instability.