Photons are promising computing medium owing to their intrinsically high bandwidth, ultrafast speed, low energy cost, low latency, and multiple orthogonal degrees of freedom enabling extreme parallelism. Photonic hardware is particularly well suited to meet artificial intelligence's enormous demand for rapidly processing compute-intensive and power-hungry workloads, thereby motivating the development of photonic neural network accelerators. Over recent decades, integrated photonic neural networks have undergone substantial architectural advances, enabling them to implement a broad range of algorithmic models, accommodate diverse data modalities, and address problems across expanding application domains. To date, both photonic-native paradigms and digitally inspired neuromorphic algorithms have been proposed and demonstrated on integrated photonic platforms. In this Review, we provide an overview of integrated photonic neural networks that emphasizes the underlying algorithms and architectures, the state-of-the-art platform implementations, and the promising application domains. Furthermore, we analyze the current challenges and provide perspectives for future developments.
Wang et al. (Sun,) studied this question.