Local Learning Rules for Out-of-Equilibrium Physical Generative Models

2025/06/23 by Cyrill Bösch, Bösch, Cyrill, Geoffrey Roeder +5 · 1 citation
Computer Science · #Emerging Technologies (cs.ET) #FOS: Computer and information sciences #FOS: Physical sciences #Machine Learning (cs.LG) #Mesoscale and Nanoscale Physics (cond-mat.mes-hall) #Neural Networks and Applications #Neural and Evolutionary Computing (cs.NE)

paper · pdf · doi:10.48550/arxiv.2506.19136

openalex publication_date 2025/06/23 · openalex created_date 2025/10/10 · openalex updated_date 2026/08/01

Abstract

We show that the out-of-equilibrium driving protocol of score-based generative models (SGMs) can be learned via local learning rules. The gradient with respect to the parameters of the driving protocol is computed directly from force measurements or from observed system dynamics. As a demonstration, we implement an SGM in a network of driven, nonlinear, overdamped oscillators coupled to a thermal bath. We first apply it to the problem of sampling from a mixture of two Gaussians in 2D. Finally, we train an oscillator network on the MNIST dataset to generate images of handwritten digits 0 and 1.

Citations

• Unsupervised and probabilistic learning with Contrastive Local Learning Networks: The Restricted Kirchhoff Machine
• Lagrangian-based Equilibrium Propagation: generalisation to arbitrary boundary conditions & equivalence with Hamiltonian Echo Learning
• Learning long range dependencies through time reversal symmetry breaking
• Nonlinear Computation with Linear Optics via Source-Position Encoding
• Programmable on-chip nonlinear photonics
• Dataset for the article 'Multimodal oscillator networks learn to solve a classification problem'
• Temporal refraction and reflection in modulated mechanical metabeams: theory and physical observation
• An analytic theory of creativity in convolutional diffusion models
• Thermodynamic Algorithms for Quadratic Programming
• Memristive Linear Algebra
• Optical Diffusion Models for Image Generation
• Training of Physical Neural Networks
• Thermodynamic Computing System for AI Applications
• Non-reciprocal topological solitons in active metamaterials
• Thermodynamic Matrix Exponentials and Thermodynamic Parallelism
• Machine learning without a processor: Emergent learning in a nonlinear analog network
• Experimental Demonstration of Coupled Learning in Elastic Networks
• Fully Non-Linear Neuromorphic Computing with Linear Wave Scattering
• Toward a formal theory for computing machines made out of whatever physics offers: extended version
• Nonlinear Processing with Linear Optics
• Realizing efficient topological temporal pumping in electrical circuits
• Training an Ising Machine with Equilibrium Propagation
• Optical Transformers
• Training precise stress patterns
• Flow Matching for Generative Modeling
• Frequency propagation: Multi-mechanism learning in nonlinear physical networks
• Photorealistic Text-to-Image Diffusion Models with Deep Language Understanding
• Score-Based Generative Modeling with Critically-Damped Langevin Diffusion
• Score-based Generative Modeling in Latent Space
• Score-Based Generative Modeling through Stochastic Differential Equations
• Denoising Diffusion Probabilistic Models
• Generative Modeling by Estimating Gradients of the Data Distribution
• Rapid Convergence of the Unadjusted Langevin Algorithm: Isoperimetry\n Suffices
• Equilibrium Propagation: Bridging the Gap Between Energy-Based Models\n and Backpropagation
• Towards Biologically Plausible Deep Learning
• Training Products of Experts by Minimizing Contrastive Divergence

Cited by

• Solving the compute crisis with physics-based ASICs

Related

• Learning to control non-equilibrium dynamics using local imperfect gradients
• Physical Derivatives: Computing policy gradients by physical forward-propagation
• Score-based generative models learn manifold-like structures with constrained mixing
• Generative Modeling by Estimating Gradients of the Data Distribution
• Micro-Macro Consistency in Multiscale Modeling: Score-Based Model Assisted Sampling of Fast/Slow Dynamical Systems