Heterogeneous nonlinear photonic spiking neuron based on a self-injection-locked laser

Neural heterogeneity, describing the diversity and complexity of biological neuron response, has been revealed to be crucial for advanced neuromorphic computing. However, in recent developments of neuromorphic photonics, the diversity of response in spiking neurons has not been fully exploited. Here, we demonstrate a heterogeneous photonic spiking neuron based on an integrated self-injection-locked (SIL) laser with tunable nonlinear responses, enabled by a distributed feedback (DFB) laser edge-coupled to a microring resonator (MRR). Mutual coupling between the laser source and MRR-based feedback structure can be used to realize neural heterogeneity, enhancing the temporal feature extraction. We use an SIL photonic spiking neuron to build a reservoir computing system with tunable nonlinearity. In experiments, it achieves 93.3% accuracy on the Iris dataset with only 80 virtual nodes. We also create three-layer heterogeneous spiking neural networks with learnable spiking activation thresholds. We demonstrate that the heterogeneous spiking neural networks achieve 21.4% improvement on temporal signal prediction tasks compared to their homogeneous counterparts, confirming the great potential of learning with heterogeneity for next-generation information processing.