Training single-electron and single-photon stochastic physical neural networks
Tong Dou, Shiro Kumara, Josh Burns, Ethan Sigler, Parth Girdhar, David Petty, Gerard Milburn, Jo Plested
Abstract
The computational demands of deep learning motivate the investigation of alternative approaches to computation. One alternative is physical neural networks~(PNNs), in which learning and inference are performed directly via physical processes. Stochastic PNNs arise when the underlying neurons are realized by the dynamics of a stochastic activation switch. Here we propose novel electronic and photonic stochastic neurons. The electronic realization is implemented by single-electron tunneling through a quantum dot. The photonic realization is implemented via a single-photon source driving one of two modes coupled via a controllable beam-splitter-like interaction. In the electronic case, the charge state of the quantum dot forms the basis for the stochastic neuron, whereas in the photonic case the occupation of the undriven mode serves as the basis for the stochastic neuron. Training of stochastic PNNs is performed with models of stochastic neurons, as well as with coherently-driven, single-photon detector stochastic neurons previously introduced. Several training strategies for MNIST handwritten digit classification have been investigated using single-hidden-layer stochastic PNNs, including varying the number of trials in each layer to control forward pass stochasticity and employing either true probability or empirical outputs in the backward pass to evaluate their influence on gradient estimation. We show that when empirical outputs are used in the backward pass, the network achieves more than 97\% test accuracy with few trials per layer. Despite the simplicity of the model architecture, high test accuracy is maintained in the presence of a high degree of noise and model uncertainty. The results demonstrate the potential of embracing stochastic PNNs for deep learning.
AI Impact Assessments
(3 models)Scientific Impact Assessment
1. Core Contribution
This paper proposes two novel physical stochastic neuron (PSN) designs—a single-electron transistor (SET) neuron based on quantum dot tunneling dynamics and a "true single-photon" (TSP) neuron based on a deterministic single-photon source coupled to a controllable beam-splitter interaction—and investigates training strategies for stochastic physical neural networks (PNNs) built from these components. The key conceptual contribution is treating intrinsic device stochasticity (shot noise, charge quantization) not as an impairment to be suppressed but as a feature to be embraced, framing physical hardware as implementing stochastic neural networks with Bernoulli-distributed activations.
The paper introduces an empirical gradient (EG) estimator that replaces the unknown true activation probability with its empirical sample mean in the backward pass, exploiting the "autonomous representation" property (where the derivative of the activation function can be expressed as a function of the activation itself). This is contrasted with the "true probability" (TP) approach requiring knowledge of the exact activation probability and straight-through (ST) estimators. The paper demonstrates >97% test accuracy on MNIST with a simple 784-400-10 architecture under severely sample-limited conditions.
2. Methodological Rigor
The physical modeling is grounded in established physics. The SET neuron derivation follows standard quantum dot master equation analysis, arriving at a sigmoid activation via the Fermi-Dirac distribution—a clean and well-known result. The TSP neuron analysis using the Fock-state master equation (FSME) framework is more technically involved and the factorization argument (Appendix A) is carefully presented, yielding an analytic expression for the b-mode occupation.
However, several aspects limit the rigor assessment:
3. Potential Impact
The paper sits at the intersection of quantum/nanoscale physics and machine learning, contributing to the growing field of physical neural networks. The potential impacts include:
The practical impact is currently limited by several factors: (a) the devices proposed are challenging to fabricate and operate (single-electron transistors require cryogenic temperatures; deterministic single-photon sources are still maturing); (b) the weight matrices are still assumed to be implemented digitally or via separate analog means—the paper focuses only on the activation function; (c) scalability to deeper networks and harder tasks is unexplored.
4. Timeliness & Relevance
The paper is timely given the surge of interest in physical neural networks (Wright et al., Nature 2022; Momeni et al., Science 2023; Kalinin et al., Nature 2025) and the growing concern about AI's energy consumption. The specific question of how to train networks where stochasticity is fundamental rather than perturbative is becoming increasingly relevant as hardware is pushed toward extreme efficiency regimes. The paper cites and builds upon very recent work (Ma et al., Nature Communications 2025), positioning itself at the current frontier.
5. Strengths & Limitations
Strengths:
Limitations:
Overall Assessment
This paper makes a modest but well-articulated contribution to the nascent field of stochastic physical neural networks. The physical neuron proposals are interesting but not yet experimentally validated, and the training methodology, while practically motivated, is evaluated only on a minimal benchmark. The EG estimator is the most transferable contribution. The paper would benefit significantly from experimental demonstration, harder benchmarks, deeper architectures, and comparison with established discrete gradient estimators.
Generated Apr 14, 2026
Comparison History (48)
Paper 1 advances a key bottleneck in microwave–optical transduction by integrating release-free optomechanical crystals with lithium niobate via micro-transfer printing and demonstrating coupling rates compatible with quantum-level operation alongside superconducting circuits. This is a concrete, experimentally validated step toward practical quantum interconnects, with strong timeliness and broad impact across quantum computing, photonics, and RF engineering. Paper 2 is conceptually interesting for neuromorphic/physical AI, but appears more model/proposal-and-simulation oriented with less immediate pathway to scalable hardware impact compared to Paper 1’s demonstrated device integration.
Paper 2 has higher impact potential: it proposes concrete single-electron and single-photon stochastic neuron implementations and evaluates training strategies with strong MNIST performance under noise/uncertainty, suggesting a plausible path toward energy-efficient, hardware-native AI. Its applications span neuromorphic computing, nanoelectronics, and integrated photonics, giving broader cross-field relevance and timeliness amid interest in post-CMOS/photonic AI accelerators. Paper 1 offers a valuable mechanistic diagnostic for barren plateaus in variational quantum algorithms, but its scope is narrower and nearer-term applicability is mainly methodological within VQA research.
Paper 2 proposes novel physical implementations of neural networks using single-electron and single-photon devices, bridging quantum physics, photonics, and deep learning. It demonstrates practical viability with >97% MNIST accuracy and robustness to noise, suggesting real-world applications in energy-efficient AI hardware. Paper 1 provides a useful diagnostic framework for understanding barren plateaus in variational quantum circuits through destructive interference, but is more incremental and narrower in scope—offering interpretive tools rather than enabling new capabilities. Paper 2's cross-disciplinary impact and hardware implications give it broader potential influence.
Paper 2 offers a rigorous complexity-theoretic and algorithmic advance: it identifies a nontrivial “intermediate” regime where free-fermion dynamics with certain non-Gaussian (magic) inputs remain classically efficiently simulable, via Pfaffian-polynomial reductions with provable additive-error guarantees matching shot-noise. This sharpens the boundary of quantum advantage and provides practical benchmarks for trapped-ion experiments and relevant quantum-chemistry subroutines (geminal-based methods), impacting quantum simulation, complexity theory, and chemistry. Paper 1 is innovative for physical stochastic neurons, but its near-term impact is more niche and less broadly foundational.
Paper 2 addresses a fundamental question in quantum computing—precisely delineating the boundary of quantum advantage—with broad implications across quantum simulation, quantum chemistry, and computational complexity. Its rigorous mathematical framework (Pfaffian reductions), practical benchmarking applications for trapped-ion experiments, and connections to quantum chemistry (geminal wavefunctions) give it wider interdisciplinary impact. Paper 1, while interesting in proposing stochastic physical neural networks, demonstrates results on a relatively simple architecture (single hidden layer, MNIST), limiting its immediate impact compared to Paper 2's foundational contributions to understanding quantum computational advantage.
Paper 2 addresses a critical and immediate challenge in quantum computing by benchmarking a state-of-the-art 98-qubit processor using an exascale supercomputer. It provides a highly anticipated quantitative boundary for quantum noise limits. While Paper 1 offers an innovative approach to low-energy physical neural networks, its impact is currently limited to proof-of-concept MNIST tasks, whereas Paper 2 delivers milestone results at the absolute edge of classical tractability with broad implications for the quantum hardware industry.
Paper 2 likely has higher impact: it delivers a timely, large-scale benchmark tying together an exascale supercomputer and a near-100-qubit QPU, providing quantitative boundaries between noise-tolerant and effectively random regimes. This has immediate, broadly relevant applications for quantum hardware validation, algorithm assessment, and HPC–QC co-design, with strong methodological rigor (reference simulations, statistical distinguishability tests). Paper 1 is novel in proposing single-electron/single-photon stochastic neurons, but remains more conceptual/model-level and narrower in near-term applicability and cross-field uptake.
Paper 2 addresses the critical computational demands of deep learning by proposing novel physical neural networks at the single-photon/electron level. Its potential to enable ultra-low-energy AI hardware provides immense real-world applicability and broad interdisciplinary impact across physics, hardware engineering, and machine learning. While Paper 1 makes solid theoretical contributions to quantum information, Paper 2's alignment with the urgent need for efficient AI architectures gives it significantly broader scientific and technological implications.
Paper 2 addresses a critical bottleneck in modern computing by proposing novel hardware for deep learning at the single-electron and single-photon level. Its interdisciplinary approach, bridging quantum physics and artificial intelligence, gives it a significantly broader potential impact across fields. While Paper 1 provides valuable fundamental bounds for quantum squeezing, Paper 2's application to physical neural networks targets an explosively growing domain with profound real-world computational implications and high relevance to the current AI hardware trajectory.
Paper 2 proposes novel physical implementations of neural networks using single-electron and single-photon stochastic neurons, bridging quantum physics with deep learning. Its interdisciplinary nature (quantum devices, photonics, machine learning), practical demonstrations on MNIST, and relevance to the urgent demand for energy-efficient AI hardware give it broader impact potential. Paper 1 proves an important but narrow no-go theorem for measurement-powered quantum engines, which is theoretically significant but has limited practical applications and appeals to a smaller research community.