Constrained Co-Design for Photonic Bayesian Neural Networks
2026-08-03 • Machine Learning
Machine Learning
AI summaryⓘ
The authors explain that traditional neural networks often give overly confident answers when faced with unclear or unfamiliar data, which is risky in important applications. Bayesian neural networks try to fix this by using probability to show uncertainty, but they can be slow and use a lot of resources. The authors study photonic Bayesian neural networks, which use light to speed up sampling, but face hardware limits like errors and range restrictions. They analyze which hardware issues can be fixed by training and which need actual hardware changes, then test their findings on several datasets, showing that their approach works well if the system's capabilities match the network's needs.
Neural networksBayesian neural networksUncertainty estimationOut-of-distribution dataPhotonic computingStochastic variational inferenceQuantizationHardware constraintsSamplingMachine learning robustness
Authors
Hendrik Borras, Xiao Wang, Bernhard Klein, Robin Janssen, Frank Brückerhoff-Plückelmann, Wolfram Pernice, Holger Fröning
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.