7–11 Dec 2026
The University of Sydney
Australia/Sydney timezone
AIP Congress 2026

Stochastic Physical Neural Networks with Single Photons and Spin Chains

Not scheduled
20m
Belinda Hutchinson Building (The University of Sydney )

Belinda Hutchinson Building

The University of Sydney

Abercrombie St & Codrington St NSW 2008
Contributed Oral AIP | Quantum Science and Technology (QST)

Description

The rapidly growing energy demands of AI have driven renewed interest in alternative approaches to computation. One alternative is physical neural networks (PNNs) in which inference and potentially training is performed directly by physical processes. Stochastic PNNs use stochasticity for computation, which is either classical (thermal) or quantum in nature. This stochasticity can dominate in low-energy systems and therefore we should account for its limitations and opportunities.

Conventional backpropagation-based training is not well-suited to PNNs because it typically requires global knowledge of the network dynamics and access to differentiable models of the physical transformations. Hence we consider backpropagation-free training using the model-free forward-forward algorithm [1], applied to a stochastic PNN consisting of layers of alternating fixed nonlinear physical activations and trainable linear transformations [2]. We present a mathematical condition for trainability and demonstrate benchmark vowel and handwritten digit classification results using spin chains to provide nonlinear physical activations.

Physics-aware training is a variant of backpropagation based using a physical forward pass and a digital backward pass. Stochastic physics-aware training is a generalization of physics-aware training that applies in the case where the neuron outputs are stochastic and therefore non-differentiable with respect to system parameters. We introduce and analyse physical stochastic neurons driven by single-photon sources [3]. This system is analysed using a Fock state master equation approach. We demonstrate high simulated test accuracy for the handwritten digit classification task with stochastic physics-aware training [4]. We then extend this notion to realising Temporal Convolutional Network architectures with trains of single-photon inputs.

[1] Momeni, A., et al. Science, 382(6676), 1297–1303 (2023).
[2] Kumara, S., et al (in preparation).
[3] Ma, SY., et al. Nat Commun 16, 359 (2025).
[4] Dou, T., et al. arXiv:2604.10861 (2026).

I am the presenting author Yes

Author

Mr Shiromal Kumara (School of Engineering and Technology, UNSW Canberra)

Co-authors

Dr David Petty (School of Engineering and Technology, UNSW Canberra) Mr Ethan Sigler (School of Engineering and Technology, UNSW Canberra) Prof. Gerard Milburn (National Quantum Computing Centre, Rutherford Appleton Laboratory, United Kingdom) Dr Jo Plested (School of Systems & Computing, UNSW Canberra) Mr Josh Burns (Sussex Centre for Quantum Technologies, University of Sussex, Brighton, United Kingdom) Prof. Matt Woolley (School of Engineering and Technology, UNSW Canberra) Dr Parth Girdhar (Department of Engineering Science, University of Oxford, United Kingdom) Mr Tong Dou (UNSW Canberra)

Presentation materials

There are no materials yet.