Probabilistic photonic computing with chaotic light

• Applicant: Professor Dr. Wolfram Hans Peter Pernice
• Subject Area: Hardware Systems and Architectures for Information Technology and Artificial Intelligence, Quantum Engineering Systems ; Communication Technology and Networks, High-Frequency Technology and Photonic Systems, Signal Processing and Machine Learning for Information Technology ; Optics, Quantum Optics and Physics of Atoms, Molecules and Plasmas
• Term: since 2025
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 543577862

Biological neural networks easily solve some of the most daunting computational problems and excel at prediction tasks on noisy and incomplete data which severely challenge conventional computing architectures. By mimicking neural and brain-like architectures, artificial neural networks (ANNs) have emerged as powerful tools for interpreting complex data distributions and making predictions. However, traditional ANNs can be seen as "point estimates" that do not inherently capture the uncertainty in predictions, do not perform well on noisy and incomplete data sets and are prone to overfitting the training data. Bayesian Neural Networks and more general probabilistic computing architectures, which integrate Bayesian inference into neural networks, offer a principled way to quantify uncertainties. In probabilistic computing, instead of having a single value as a parameter, the underlying networks operate on distributions over possible parameter values. This inherently requires a computationally intensive approach to model training and, in particular, for inference due to frequent sampling operations performed in each network layer. In this project, I aim to realize parallel probabilistic processors to address these challenges. The primary research question centers on whether inherent hardware noise can be harnessed for probabilistic computations. The aim is to transform a major drawback (noise potentially impacting computation accuracy) into a strength (reducing the cost of probabilistic computations by managing uncertainty). This concept will be exploited for state estimation, stochastic variational inference, Bayesian and Gaussian processing in hybrid probabilistic-deterministic neural networks.

• DFG Programme: Reinhart Koselleck Projects