Design and optimization of multi-operand ring resonator based efficient optical recurrent neural network

The rapid evolution of artificial intelligence (AI) technologies results in ascending demand for computation with high performance, low latency, and low power consumption. As Moore’s Law is approaching its physical limit, optical computing is regarded as a promising counterpart of electronic architecture for accelerating such highly intensive deep-learning-related tasks. Moreover, the development of nanophotonic technologies and photonic integrated circuits further boosts these optical neural networks (ONNs) and endows the deployment of optical computing in real-world applications with more realistic values. Among different kinds of neural networks, recurrent neural networks (RNNs) are specially designed for processing sequence information, and they predominate speech and language application scenarios. Even though many dedicated optical neural architectures have been proposed in recent years, most of them focus on feedforward networks whereas optical recurrent neural networks are less frequently investigated. In this thesis, compact and efficient Elman-like optical RNN model and Gated Recurrent Unit (GRU)-like ONN model are implemented by utilizing multi-operand ring resonator (MORR) and dense wavelength division multiplexing (DWDM). Network compression schemes and specialized training optimizations are introduced to enhance the efficiency and robustness of the architecture. Optical simulations are conducted to verify the fidelity of this optical recurrent neural network, and experiments of applying it in real-world applications prove the high accuracy, small footprint, and high efficiency of the proposed design.Electrical and Computer Engineerin