Constrained Co-Design for Photonic Bayesian Neural Networks

Research Context and Approach

Bayesian neural networks (BNNs) are essential for safety-critical AI because they provide uncertainty estimates, but their reliance on repeated sampling makes them computationally expensive on digital hardware. Photonic probabilistic computing offers a high-speed alternative by leveraging intrinsic optical stochasticity. However, photonic hardware is not an ideal sampler; it imposes physical constraints—such as quantization, limited dynamic range, and programming errors—that restrict the variational distributions the network can represent. This paper treats the photonic processor as a constrained stochastic operator, systematically evaluating how these hardware limitations affect uncertainty quality and predictive performance.

Co-Design Methodology

The researchers formulate photonic BNN inference as a constrained stochastic variational inference problem. They define four primary stochastic operator variants based on the location (weight vs. activation) and modality (additive vs. multiplicative) of the noise. By implementing these operators in a probabilistic programming framework, the team performed an ablation study to determine the sensitivity of the network to various hardware constraints. This approach allows for the classification of constraints into those that can be compensated for during training (e.g., via hardware-aware training) and those that necessitate fundamental hardware or architectural modifications.

Key Findings and Guidelines

The study demonstrates that hardware-aware training can effectively recover predictive performance and uncertainty quality, provided the required variational family remains representable within the hardware's constraints. The authors provide concrete guidelines for designers, identifying specific thresholds for quantization, mean/variance bounds, and programming noise tolerance. These guidelines help distinguish between benign hardware limitations and those that fundamentally degrade the model's ability to quantify uncertainty, offering a roadmap for scaling photonic BNNs beyond small-scale proof-of-concept demonstrations.