Constrained Co-Design for Photonic Bayesian Neural Networks

2026-08-03

AI Summary

This work addresses the challenge of deploying Bayesian neural networks (BNNs) on photonic hardware, where analog constraints—such as quantization, programming errors, and limited dynamic range—hinder efficient uncertainty-aware inference. The authors formulate photonic BNN inference as constrained stochastic variational inference and systematically analyze how hardware limitations affect the expressiveness of the variational posterior family. They propose a hardware-software co-design principle that distinguishes between constraints compensable through training and those requiring hardware-level mitigation. By incorporating mean/variance bound constraints, explicit error modeling, and ablation of stochasticity placement, they demonstrate on Dirty-MNIST, CIFAR-10, and CINIC-10 that hardware-aware training recovers both predictive accuracy and uncertainty quality without sacrificing distributional expressiveness. This study presents the first large-scale BNN inference under realistic photonic hardware constraints.

Abstract

Classical neural networks frequently produce overconfident predictions on ambiguous or out-of-distribution (OOD) data, a liability that grows with each AI system deployed in safety-critical real-world scenarios. Bayesian neural networks (BNNs) provide a principled framework for uncertainty-aware prediction by replacing deterministic parameters with probability distributions, but repeated sampling increases latency, memory traffic, and energy consumption. Photonic probabilistic computing offers a promising alternative by exploiting intrinsic optical stochasticity for fast and parallel sampling. However, photonic BNNs are not ideal samplers: analog constraints on quantization, programming error, dynamic range, and representable mean and variance restrict the variational families that can be implemented in hardware. In this work, we study which hardware-imposed constraints limit scalable photonic BNN inference, how these constraints can be represented, and which ranges can be tolerated by photonic BNNs beyond small proof-of-concept networks. We formulate photonic BNN inference as constrained stochastic variational inference and perform a systematic ablation study over stochasticity location, stochasticity modality, quantization, programming error, and mean/variance bounds. From these results, we derive concrete co-design guidelines that distinguish hardware constraints that can be compensated by training from those requiring hardware or architecture intervention. We validate these guidelines under coupled, hardware-realistic constraints on Dirty-MNIST, CIFAR-10, and CINIC-10, using Fashion-MNIST and SVHN as OOD benchmarks, showing that hardware-aware training recovers predictive performance and uncertainty quality whenever the required variational family remains representable, whereas violations of representational limits require targeted hardware modifications.

Problem

• Photonic Bayesian Neural Networks
• Hardware Constraints
• Uncertainty Quantification
• Variational Inference
• Out-of-Distribution Detection

Innovation

• photonic Bayesian neural networks
• constrained stochastic variational inference
• hardware-aware co-design
• uncertainty quantification
• analog hardware constraints