Learning without neurons in physical systems

What to take away

1. 1. Local learning rules of the form dw[x,t]/dt ∼ h(s(f;{w})[x,t]) — where a learning degree of freedom changes only based on the state of the system in its own spacetime vicinity — are sufficient to collectively minimize a global cost function because the physical response encodes global information locally.
2. 2. Contrastive learning in an experimental resistor network (Dillavou et al. 2021, arXiv:2108.00275) classifies the 4-feature, 3-class Iris dataset, achieving accuracy that rises from ~83% at step 1 to ~97% at step 300, demonstrating supervised physical learning without any computer in the learning loop.
3. 3. A Hebbian molecular learning rule dwij/dt ∼ si(x,t)sj(x,t), implemented via proximity-based ligation, enables a 2,500-molecular-species system to perform pattern recognition on 2,500-pixel images through stochastic non-equilibrium nucleation dynamics (Zhong, Schwab, Murugan 2017, J. Stat. Phys. 167:806).
4. 4. Learning capacity in 2D molecular self-assembly systems scales as the square root of the number of distinct molecular species (Murugan et al. 2015, PNAS 112:54), so doubling the number of species increases capacity by a factor of ~1.41.
5. 5. Even when the learning timescale τlearn is comparable to the physical response timescale τresponse (ratio R = 1), contrastive learning in flow networks still converges successfully; only at rates significantly exceeding R = 1 do underdamped oscillations emerge that may prevent convergence (Stern et al. 2022, Phys. Rev. Research 4:L022037).
6. 6. Trained physical systems develop exponentially fewer energy minima than random disordered systems of the same size; in learned molecular self-assembly, spurious chimeric minima near capacity are structurally analogous to spurious states in Hopfield associative memory (Hopfield 1982, PNAS 79:2554).
7. 7. Supervised 'thumbs-up/thumbs-down' learning in disordered creased sheets — where the supervisor selects the sign of the unsupervised rule by placing the sheet in one of two chemical environments — is sufficient to classify subtle spatial force-pattern correlations mapped from the Iris dataset (Stern et al. 2020, PNAS 117:14843).
8. 8. To replicate contrastive learning experimentally, a researcher can couple two copies of the same resistor network (a 'free' and a 'clamped' network), update each variable resistor Ri by ΔRi = +δR if |ΔViC| > |ΔViF| and −δR otherwise, and drive the system with voltage inputs encoding Iris dataset features, as demonstrated in Wycoff et al. 2022 (J. Chem. Phys. 156:144903).
9. 9. An open question raised by the paper is whether non-equilibrium physical systems — those without time-reversal symmetry, such as molecular motor-driven cytoskeletal networks — are intrinsically more expressive or show greater parameter degeneracy for a given behavior, which could make them easier to train than equilibrium systems.
10. 10. Physical learning produces a diagnostic signature of reduced response dimensionality: trained elastic and flow networks develop soft modes (low-eigenvalue normal modes) so that responses to arbitrary random forces are confined to a low-dimensional subspace, a feature absent in randomly parameterized networks of the same topology (Tlusty, Libchaber, Eckmann 2017, Phys. Rev. X 7:021037).

Peer brief — for seminar discussion

Stern and Murugan's 2022 review consolidates and theoretically frames an emerging program in which physical materials — disordered elastic networks, flow networks, creased sheets, and molecular assemblies — autonomously solve inverse problems by experiencing examples of desired behavior, modifying their own parameters through local dynamical rules, with no computer involved in the adaptation loop. The central organizing concept, which they call physical learning, is formalized around three ingredients: physical degrees of freedom that map inputs to states, learning degrees of freedom (bond stiffnesses, tube radii, molecular interaction strengths) that modulate those mappings, and a learning rule dw[x,t]/dt ∼ h(s(f;{w})[x,t]) that updates each learning element based only on the local state of the system at that spacetime point. The load-bearing theoretical claim is that such purely local rules can nevertheless minimize a global cost function, because the physical response of the system encodes global information in local state variables through collective dynamics — no explicit gradient computation is required. The most quantitatively grounded demonstrations span several substrates. A contrastive learning rule implemented in an experimental resistor network (Dillavou et al. 2021) classifies the 4-feature Iris dataset, improving from roughly 83% accuracy at training step 1 to 97% at step 300. A Hebbian molecular rule (dwij/dt ∼ si·sj) applied to a 2,500-species chemical system performs pattern recognition on 2,500-pixel images via stochastic nucleation (Zhong et al. 2017, J. Stat. Phys. 167:806). Learning capacity in 2D molecular self-assembly scales as the square root of the number of distinct species (Murugan et al. 2015, PNAS 112:54). Separate work shows that contrastive learning in flow networks converges even when the learning timescale matches the physical response timescale (Stern et al. 2022, Phys. Rev. Research 4:L022037), relaxing a constraint that in silico ML takes for granted. The framework introduced to unify these results is the classification of physical learning rules into unsupervised (Hebbian, flux-based, directed aging) and supervised (thumbs-up/thumbs-down, contrastive Hebbian) categories, with the contrastive rule identified as the most powerful physically realizable approach to supervised learning. An alternative method the review explicitly sets aside is computer-aided backpropagation as implemented in Wright et al. 2022 (Nature 601:549), which uses physical systems as substrates but retains a computer in the weight-update loop; the review treats this as a related but distinct program. The paper also identifies physical signatures that learning leaves in its substrate — spatial heterogeneity in moduli, topological pruning of unused edges, anomalously few energy minima relative to random disordered systems, and the emergence of soft modes that confine responses to a low-dimensional subspace — and proposes these as empirical markers for detecting learning or evolution in natural physical systems. The prediction the review advances is that non-equilibrium physical systems lacking time-reversal symmetry may prove more expressive than equilibrium ones, because non-equilibrium nucleation dynamics already demonstrably improve pattern-recognition capacity, and non-reciprocal interactions could unlock dynamic-phase learning not accessible at steady state. A critical reader would push back on the scope conflation between genuinely autonomous physical learning and systems that still require substantial human supervision to implement the learning rule. The 'thumbs-up/thumbs-down' protocol in creased sheets requires a human supervisor to decide the sign of the rule for each training example and to physically place the material in different chemical environments accordingly; the contrastive rule in resistor networks requires constructing two coupled copies of the network and engineering the update circuit. These are meaningful engineering achievements, but calling them 'physical learning without computers' overstates autonomy. The degree to which the training protocol itself can be internalized by the physical system — rather than externally administered — remains largely unresolved experimentally, and the review's framing tends to blur this distinction.

Methods (3)

• Local learning rule
  Learning rule where change in a parameter at point x,t depends only on system state at same or nearby spacetime points, without requiring global cost function computation
• Molecular Hebbian learning
  Unsupervised learning rule in molecular systems where species i,j with high co-localized concentrations strengthen their interaction strength through proximity-based ligation
• Spike-Timing-Dependent Plasticity
  Biologically plausible local learning rule constraining the brain; referenced as precedent for locality-constrained learning in physical systems

Frameworks (4)

• Contrastive learning
  Supervised learning framework where system learns by observing contrast between current response and nudged improved response; requires weak additional forces from supervisor
• Neuromorphic computing
  Related field where physical elements are modified for desired computational ability; traditionally targets symbolic inputs/outputs unlike physical learning's physical stimuli/responses
• Physical learning
  Framework for solving inverse problems in which physical systems autonomously adapt their parameters in response to stimuli through local learning rules, without requiring computational design or explicit cost functions
• Reservoir Computing
  Physical computation framework using fixed complex dynamical systems with trained output filter; contrasted with physical learning's parameter modification approach

Findings (1)

• Folding pathways of creased sheets can be trained for specific topologies including classification of mechanical force patterns analogous to neural networks

  Experimentally validated finding that origami/kirigami systems can solve classification tasks through physical learning of crease stiffnesses

Claims (2)

• Physical systems are more constrained in learning abilities than in silico neural networks due to locality requirements, but this mirrors biological learning constraints and offers robustness benefits

  Core theoretical claim establishing that locality constraints in physical learning are not fatal—they reflect biological precedent and provide advantages like robustness and scalability

• The 'muscle' itself—peripheral biological tissue—can learn and make decisions autonomously, not just centralized control units like the brain

  Theoretical claim that physical learning reveals non-modular information processing; contrasts traditional view of separated control (brain) from controlled elements (muscle)

Hypotheses (1)

• Collective nucleation dynamics in Hebbian-learned molecular interaction systems can perform pattern recognition by assembling different structures in response to different concentration patterns

  Theoretical prediction that molecular systems with proximity-based learning can recognize patterns; has mathematical connections to Hopfield associative memory

Questions (1)

• To what extent can physical systems learn by exploiting typically local natural processes without any explicit cost function?

  Central research question defining the scope of physical learning; asks about achievable learning under locality constraints

Related work— refs + corpus + external arXiv

Cited / in-corpus / arXiv badges show which signals surfaced each row. Multi-source rows weighted higher.

• Supervised learning through physical changes in a mechanical system

  cited

  2020
  ≈ 87%
• Learning Biomolecular Motion: The Physics-Informed Machine Learning Paradigm

  Aaryesh Deshpande

  2025
  ≈ 85%
• Competitive plasticity to reduce the energetic costs of learning

  Mark CW van Rossum

  2026
  ≈ 84%
• The Physical Basis of Prediction: World Model Formation in Neural Organoids via an LLM-Generated Curriculum

  Brennen Hill

  2025
  ≈ 84%
• Energy-Based Dynamical Models for Neurocomputation, Learning, and Optimization

  Francesco Bullo, Dmitry Krotov, Adilson E. Motter Arthur N. Montanari

  2026
  ≈ 84%
• Beyond spiking networks: the computational advantages of dendritic amplification and input segregation

  Cosimo Lupo, Paolo Muratore, Pier Stanislao Paolucci Cristiano Capone

  2026
  ≈ 84%
• Improved Learning in Evolution Strategies via Sparser Inter-Agent Network Topologies

  Dan Calacci, Yan Leng, Peter Krafft, Esteban Moro, Alex Pentland Dhaval Adjodah

  2019
  ≈ 83%
• Predicting Brain Morphogenesis via Physics-Transfer Learning

  Yingjie Zhao and Yicheng Song and Fan Xu and Zhiping Xu

  2025
  ≈ 83%
• A unified theory of learning

  Taisuke Katayose

  2022
  ≈ 83%
• Deep Neuroevolution of Recurrent and Discrete World Models

  Kenneth O. Stanley Sebastian Risi

  2019
  ≈ 83%
• An effective theory of collective deep learning

  Llu\'is Arola-Fern\'andez and Lucas Lacasa

  2023
  ≈ 83%
• Model-based Reinforcement Learning with a Hamiltonian Canonical ODE Network

  Yuhong Jiang, Hang Su, Dong Yan, Jun Zhu Yao Feng

  2022
  ≈ 83%
• Neural Operator: Is data all you need to model the world? An insight into the paradigm of data-driven scientific ML

  Md Ashiqur Rahman, Abhijeet Vyas, Andrey Shor, Beatriz Medeiros, Stephanie Hernandez, Suhas Eswarappa Prameela, Aniket Bera Hrishikesh Viswanath

  2026
  ≈ 83%
• What Neuroscience Can Teach AI About Learning in Continuously Changing Environments

  Bruno Averbeck, Georgia Koppe Daniel Durstewitz

  2025
  ≈ 83%
• Differential learning kinetics govern the transition from memorization to generalization during in-context learning

  Gautam Reddy Alex Nguyen

  2024
  ≈ 83%
• Learning Developmental Scaffoldings to Guide Self-Organisation

  Milton L. Montero and Elias Najarro and Jakob Schauser and Sebastian Risi

  2026
  ≈ 83%
• Active inference on discrete state-spaces: a synthesis

  in corpus

  2020
  ≈ 83%
• Technological Approach to Mind Everywhere: An Experimentally-Grounded Framework for Understanding Diverse Bodies and Minds

  in corpus

  2022
  ≈ 82%
• The computational boundary of a 'self': developmental bioelectricity drives multicellularity and scale-free cognition

  in corpus

  2019
  ≈ 82%
• Topological constraints on self-organization in locally interacting systems

  in corpus

  2026
  ≈ 82%
• Multiple ways to implement and infer sentience

  in corpus
  ≈ 82%
• Differentiable Logic Cellular Automata: From Game of Life to pattern generation with learned recurrent circuits

  in corpus
  ≈ 81%
• Generalizing frameworks for sentience beyond natural species

  in corpus
  ≈ 81%
• Developmental Bioelectricity: the cognitive glue enabling evolutionary scaling from physiology to mind

  in corpus

  2023
  ≈ 81%
• Self-Improvising Memory: A Perspective on Memories as Agential, Dynamically Reinterpreting Cognitive Glue

  in corpus

  2024
  ≈ 81%
• Collective intelligence: A unifying concept for integrating biology across scales and substrates

  in corpus

  2024
  ≈ 81%
• The biogenic approach to cognition

  in corpus

  2005
  ≈ 80%
• The Platonic Representation Hypothesis

  in corpus

  2024
  ≈ 80%
• A Free energy principle for the brain (lecture summary)

  in corpus

  2008
  ≈ 80%
• The collective intelligence of evolution and development

  in corpus

  2023
  ≈ 80%

+28 more

Similar preprints — Semantic Scholar

Cited by (1)

• The collective intelligence of evolution and development

  Watson and Levin argue that evolutionary individuality, organismic individuality, and cognition are coextensive — the causal structures necessary to produce fitness that belongs to a collective rather