Training of Physical Neural Networks

2024/06/05 by Ali Momeni, Momeni, Ali, Babak Rahmani +55 · 4 voices · 43 citations
Computer Science · Materials Science · Physics and Astronomy · #Artificial intelligence #Artificial neural network #Backpropagation #Computer science #Data science #Inference #Leverage (statistics) #Machine Learning in Materials Science #Machine learning #Neural Networks and Applications #Scale (ratio) #cs.LG #physics.app-ph

paper · pdf · doi:10.48550/arxiv.2406.03372

published in arXiv (Cornell University) (Cornell University)

openalex publication_date 2024/06/05 · openalex created_date 2025/09/03 · openalex updated_date 2026/08/01

Abstract

Physical neural networks (PNNs) are a class of neural-like networks that leverage the properties of physical systems to perform computation. While PNNs are so far a niche research area with small-scale laboratory demonstrations, they are arguably one of the most underappreciated important opportunities in modern AI. Could we train AI models 1000x larger than current ones? Could we do this and also have them perform inference locally and privately on edge devices, such as smartphones or sensors? Research over the past few years has shown that the answer to all these questions is likely "yes, with enough research": PNNs could one day radically change what is possible and practical for AI systems. To do this will however require rethinking both how AI models work, and how they are trained - primarily by considering the problems through the constraints of the underlying hardware physics. To train PNNs at large scale, many methods including backpropagation-based and backpropagation-free approaches are now being explored. These methods have various trade-offs, and so far no method has been shown to scale to the same scale and performance as the backpropagation algorithm widely used in deep learning today. However, this is rapidly changing, and a diverse ecosystem of training techniques provides clues for how PNNs may one day be utilized to create both more efficient realizations of current-scale AI models, and to enable unprecedented-scale models.

Cited by

• Multi-layer MIMO Relay as Deep Physical Neural Networks: Power Amplifiers as Activation Functions
• On the Information Required for Feedback Control
• Towards transferable lightweight neuromorphic computing through a model-free temporal-switch framework
• Physical Reservoir Signal Acquisition for Sub-Nyquist Waveform Reconstruction
• High-capacity associative memory in a quantum-optical spin glass
• Self-configuring high-speed multi-plane light conversion
• Stacked Intelligent Metasurface-Aided Wave-Domain Signal Processing: From Communications to Sensing and Computing
• Optical kernel machine with programmable nonlinearity
• Analog Physical Systems Can Exhibit Double Descent
• Spectroscopy on a single nonlinear mode recognizes quantum states
• Miniaturised transmissive multi-plane light converters via laser-written geometric phase holograms cascaded in glass
• Tutorial: A practical guide to the alignment of defocused spatial light modulators for fast diffractive neural networks
• Towards Engineering Material Neural Networks
• Neural networks for neurocomputing circuits: a computational study of tolerance to noise and activation function non-uniformity when machine learning materials properties
• MeSH: Memory-as-State-Highways for Recursive Transformers
• Optical Computing with Spectrally Multiplexed Features in Complex Media
• Intelligent Mode Sorting in Turbulence with Task-Dependent Optical Neural Networks
• Training deep physical neural networks with local physical information bottleneck
• Unsupervised and probabilistic learning with Contrastive Local Learning Networks: The Restricted Kirchhoff Machine
• Programmable k-local Ising Machines and all-optical Kolmogorov-Arnold Networks on Photonic Platforms
• Low-rank surrogate modeling and stochastic zero-order optimization for training of neural networks with black-box layers
• Thermodynamic uncertainty relation for generalized time-reversal observables
• Self-Organising Memristive Networks as Physical Learning Systems
• On the possibility of deep alignment
• Learning with springs and sticks
• Hybrid Deep Reconstruction for Vignetting-Free Upconversion Imaging through Scattering in ENZ Materials
• Higher-Order Kuramoto Oscillator Network for Dense Associative Memory
• Beyond Silicon: Materials, Mechanisms, and Methods for Physical Neural Computing
• Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation
• Solving the compute crisis with physics-based ASICs
• Model-free Optical Processors using In Situ Reinforcement Learning with Proximal Policy Optimization
• Towards the Training of Deeper Predictive Coding Neural Networks
• Equilibrium Propagation for Dissipative Dynamics
• Local Learning Rules for Out-of-Equilibrium Physical Generative Models
• Machine Intelligence on Wireless Edge Networks
• Inverse-designed nanophotonic neural network accelerators for ultra-compact optical computing
• Learning long range dependencies through time reversal symmetry breaking
• Mechanical in-sensor computing: a programmable meta-sensor for structural damage classification without external electronic power
• Programmable metasurfaces for future photonic artificial intelligence
• Scalable Photonic Neural Networks via Surrogate Scattering-Matrix Inverse Design
• Background cancellation for frequency-selective quantum sensing
• Passive All-Optical Nonlinear Neuron Activation via PPLN Nanophotonic Waveguides
• Artificial neuron [wikipedia]

Discussions

• Training of Physical Neural Networks [hn, 142 points, 46 comments]
• #PhysicsJournalClub "Training of Physical Neural Networks" by a LOT of people (including @FMarquardtGroup and @sylvaingigan, who I think are the only ones with a presence here on bsky) arxiv.org/pdf/2 [bsky, 6 points, 1 comments]
• Training of Physical Neural Networks (arxiv.org) Main Link | Discussion [bsky, 1 points, 0 comments]
• Physical이 실제 생명체의, 특히 인간의 그것이라는 말은 아니고, 전기소자를 이용한 물리적 인공신경망의 훈련에 대한 연구. 이론적으로는 현재의 디지털 연산을 이용한 구현체보다 훨씬 더 빠르고 효율적. Training of Physical Neural Network. arxiv.org/abs/2406.03372 [bsky, 0 points, 0 comments]

Related

• Perfecting Imperfect Physical Neural Networks with Transferable Robustness using Sharpness-Aware Training
• Training deep physical neural networks with local physical information bottleneck
• Backpropagation-free Training of Deep Physical Neural Networks
• Physical Deep Learning with Biologically Plausible Training Method
• Physics-informed ConvNet: Learning Physical Field from a Shallow Neural Network