UNSW's Quantum Neuron Blueprint Trains Neural Networks on Single Electrons and Photons
Researchers propose single-electron and single-photon stochastic neurons, then train them with sample-based gradient estimators reaching 97 percent on MNIST.
ProTakeaways
- New paper proposes single-electron transistor and true single-photon stochastic neurons for physical neural networks. arxiv
- Both neurons produce binary outputs from single quantum events, with tunable activation probability.
- Empirical gradient estimator trains networks using only sampled outputs, not analytic activation probabilities.
- MNIST accuracy exceeds 97 percent with just 5 to 10 samples per neuron per forward pass.
- Identity straight-through estimator saturates at 93 percent, showing estimator choice matters significantly.
- Framework extends physics-aware training from Ma et al. 2025 to quantum-limited electronic hardware.
A team from UNSW Canberra, Sussex, Oxford, and the UK's National Quantum Computing Centre has laid out a blueprint for neural networks whose neurons are individual quantum events, either a single electron tunneling through a quantum dot or a single photon hopping between two optical modes. Their paper shows these devices can be trained on MNIST to more than 97 percent test accuracy using only a handful of noisy binary samples per neuron.
The work sits inside a growing effort to build physical neural networks (PNNs), where the arithmetic of a forward pass is not simulated on a GPU but performed directly by the physics of an engineered device. Wright et al. (2022) demonstrated the idea by training everything from optical to mechanical systems with backpropagation. Push these devices toward extreme energy efficiency and the signals become discrete and probabilistic, at which point standard training assumptions break down.
When noise stops being noise
Physical neural networks perform learning and inference directly via physical processes, and stochastic PNNs arise when the underlying neurons are realized by the dynamics of a stochastic activation switch. Most analog hardware papers treat device noise as a small perturbation and use noise-aware backpropagation to compensate. That framing collapses in the regime this paper targets: when a neuron activation is triggered by a single photon detection or a single electron tunneling event, the output is a coin flip rather than a slightly noisy real number.
The precedent here is Ma et al. (2025) from the McMahon group, who experimentally demonstrated MNIST classification at 98 percent test accuracy using an optical neural network operating in the single-photon regime, corresponding to just 0.038 photons per multiply-accumulate operation. Their trick, called physics-aware stochastic training, was to bake the probabilistic model of the single-photon detector directly into the training loop. The new paper generalizes and stress-tests that idea across three different physical neuron designs.