Photonic neural networks (PNNs) represent an important class of optical computing with the goal of producing an accelerated processor that combines the information processing capacity of neu- romorphic systems, and the speed and bandwidth of photonics. This thesis focuses on system de- sign, experimental demonstration and AI applications of PNNs using integrated photonics. Two main thrusts of the PNNs development in this thesis are: studying bio-inspired spiking network on InP-based integrated photonic circuits, and building scalable continuous-time neural network using silicon photonics.
Toward the first thrust, we study the temporal dynamics of an integrated excitable laser, and demonstrate its analogy to a biological spiking neuron and its compatibility for large-scale system integration. With a solid experimental demonstration, we further propose the model of such photonic spiking neural network, and show its applications including temporal XOR task, time series processing, and recommendation systems. For the second thrust, we investigate a silicon photonics-based system to achieve both precise weight control and programmable nonlinearity.
We further explore its application to real-world problems in communication systems. The pro- posed compact model using silicon photonic recurrent neural network enables real-time specific emitter identification, and provides a promising platform for future edge AI systems.