Introducing probabilistic models into photonic neural networks, harnessing high-throughput and low-latency performance of photons, holds great promise for Bayesian inference. Photonic spiking neurons with stochasticity are essential to realize probabilistic computing. However, existing probabilistic neurons lack intrinsic stochasticity and rely on external entropy sources, adding architectural complexity that impedes high-density integration. Here, we report the first compact on-chip photonic spiking neuron with inherent stochasticity based on a novel phase-change material SbTe9, featuring an active footprint of only 1.5 μm2. This neuron enables stable and tunable probabilistic firing behaviors, arising from the intrinsic fluctuations of the melting point and temperature of the SbTe9 layer driven by microstructure evolution during non-equilibrium melting and crystallization. Leveraging this stochasticity, the neuron enables the Bayesian inference achieving 98.67% accuracy with uncertainty quantification for breast cell diagnosis, and demonstrates remarkable tolerance to hardware synaptic variations (0.47% reduction, ten times smaller) and input noise (4.28% reduction at 15% noise, over twofold smaller) compared with deterministic neurons. Based on the novel volatile phase-change material, this neuron establishes a transformative pathway toward the development of large-scale, low-complexity and high-performance on-chip photonic neuromorphic computing systems. Efficient probabilistic computing requires hardware with inherent stochasticity. Here, the authors demonstrate a compact on-chip photonic neuron utilizing a novel SbTe9 phase-change material to achieve stable, tunable probabilistic firing for Bayesian inference.
董云肖 et al. (Mon,) studied this question.