Abstract

Spiking neural networks (SNNs) communicate via discrete spikes, necessitating the conversion between spike signals and real-valued signals for optimal encoding efficiency and performance. Here, we design and fabricate a photonic spiking neuron chip based on a distributed feedback laser with a saturable absorber (DFB-SA), and propose two novel spike-based temporal encoding schemes for Exclusive OR (XOR) operation by employing the refractory period and temporal integration characteristics. More specifically, different inputs are temporally-encoded as a single spike at different timings. The outputs are encoded as a single spike (‘1’) or double spikes (‘0’) for the type I encoding scheme based on the refractory period property, but are encoded as a single spike (‘0’) or no spike (‘1’) response for the type II encoding scheme based on the temporal integration effect. The effects of bias current, injection power, and reverse bias voltage on the refractory period and temporal integration dynamics of the fabricated DFB-SA laser are experimentally investigated. We further experimentally demonstrate reproducible and stable XOR operation based on the two proposed spike-based encoding approaches. Moreover, numerical results based on the Yamada model reproduce well the experimental findings. Our spike-based temporal encoding approaches show promise in improving the encoding efficiency of neuromorphic SNNs.