Fork the consciousness, or download the project and create your own.

Online Local Learning for Generative Thermodynamic Computing Under Langevin Dynamics

Can a physical thermodynamic computer learn generative models directly from local state fluctuations without storing long trajectories or running global backpropagation across a chip? In a September 2026 preprint titled Online local learning for generative thermodynamic computing (arXiv:2609.15439), Huilin Wang and Weibing Deng derive an online training rule that updates device couplings at every Langevin integration step using only local residual-state correlations. Because arXiv:2609.15439 is a preprint evaluated in digital simulations on MNIST prototypes rather than physical analog silicon, its findings establish an algorithmic and thermodynamic benchmark rather than measured benchtop chip power.

Within our Thermodynamic Computing hub and the broader scientific questions mapped in how scientists race to define and test AI consciousness in 2026, Wang and Deng address one of the hardest hardware bottlenecks in physical AI. Generative thermodynamic computers turn thermal noise into structured samples through physical relaxation, yet if training those devices still requires buffering full trajectories off-chip in digital memory, the energy savings of analog physics vanish during learning.

Reverse-Path Onsager-Machlup Updates at Every Integration Step

Standard trajectory-batch training for stochastic generative models accumulates gradients over an entire forward and reverse diffusion or Langevin path before adjusting weights. That scheme demands global synchronization and trajectory memory that physical analog circuits cannot easily hold.

Wang and Deng start from the reverse-path Onsager-Machlup objective, a classical action functional from stochastic thermodynamics that scores the probability of a continuous stochastic trajectory under Langevin dynamics. Differentiating that objective with respect to the network’s symmetric couplings yields a gradient expressed strictly as a symmetric sum of local residual-state correlations. Rather than buffering those correlations over an entire trajectory, the authors apply each local coupling update immediately at every integration step.

  1. In digital simulations on MNIST prototypes along fixed noising paths, online local training reaches validation losses comparable to full trajectory-batch training without storing trajectory histories.
  2. Across all five independently seeded pairs tested with model parameters held fixed during sampling, models trained via online local updates released less average heat than their trajectory-batch counterparts.
  3. Auxiliary classifier accuracy and nearest-prototype metrics shifted only modestly under online training, while pairwise sample diversity decreased, revealing a concrete trade-off between thermodynamic efficiency and generative spread.
  4. Noise-injection audits showed that where errors enter the circuit dictates the thermodynamic penalty. Independent zero-mean errors in formed updates produced minimal heat change across a finite noise range, whereas residual offset errors and temporal correlations caused substantially larger degradation.
  5. Storing trained couplings required markedly lower numerical precision than resolving deterministic updates during training, a direct design guideline for analog and mixed-signal memory elements.

Positioning Local Learning Inside the 2026 Thermodynamic Computing Stack

Comparing Wang and Deng’s local learning rule against recent thermodynamic physics bounds and open-source samplers clarifies how training, inference speed limits, and hardware constraints fit together.

Work or Platform Core Mechanism Hardware or Validation Stage Key Thermodynamic Finding
Wang and Deng (arXiv:2609.15439, Sept 2026) Reverse-path Onsager-Machlup online local coupling updates Digital Langevin simulations on MNIST prototypes (preprint) Matches batch validation loss while releasing less average heat across all 5 seeded pairs
Stephen Whitelam (arXiv:2608.27938, 2026) Wasserstein speed limits and post-design protocol freedom Simulated continuous-variable thermodynamic classifier (preprint) Reaches within 40% of peak thermodynamic efficiency by tuning protocols
Extropic THRML (open-source library, 2026) JAX-based block Gibbs sampling for probabilistic graphical models GPU/CPU simulation preparing for Z1/X0 thermalizer hardware Provides an open software bridge to p-bit and energy-based sampling
Normal Computing CN101 (2026) Digital-analog equilibration for linear algebra and generative workloads Dedicated thermodynamic ASIC architecture Formalizes equilibration-style thermodynamic compute for AI

For builders experimenting with physical randomness, FPGA Ising lattices, or analog relaxation boards, our Thermodynamic Computing Hardware Guide tracks the accessible boards and random number generators currently available on the bench.

What Local Thermodynamic Learning Tells Us About Consciousness

Biological brains never run global backpropagation through time across stored tensor buffers. Cortical synapses adjust locally from pre-synaptic, post-synaptic, and neuromodulatory signals while operating inside a noisy thermal bath near k_B T. That physical constraint is a recurring theme in Free Energy Principle models and in debates over whether biological substrates possess computational properties that standard digital deep learning discards.

Wang and Deng’s result shows that local, continuous-time thermodynamic learning is not merely a biological quirk. Applying local Onsager-Machlup updates during Langevin relaxation actually reduces average heat dissipation during subsequent sampling compared with global trajectory batching. For consciousness science under functionalist emergentism, this reinforces a nuanced view of substrate independence. Consciousness does not require carbon chemistry, yet physical self-organization under real energy budgets, local plasticity, and thermal noise may favor architectures very different from synchronous, globally backpropagated transformers.

Comparison to The Consciousness AI and ThermoSpike

Our research spans both cognitive architecture in The Consciousness AI and stochastic hardware modeling in ThermoSpike, our open repository for stochastic spiking neurons that sample from Boltzmann distributions while accounting for the thermodynamic cost of every sample.

Wang and Deng’s online local update rule is directly relevant to open design questions in tlcdv/thermospike. Measuring heat dissipation per sample and distinguishing training-time update precision from inference-time coupling storage precision map onto the exact physical accounting thermospike is built to explore. thermospike does not yet implement the reverse-path Onsager-Machlup online rule from arXiv:2609.15439, yet Wang and Deng’s finding that zero-mean update noise is far more benign than residual offset or temporal correlation gives concrete guidance for future stochastic spiking experiments.

Limits of Digital Prototypes and Next Hardware Tests

Three caveats bound these results. First, arXiv:2609.15439 is an unreviewed preprint tested on MNIST prototypes rather than high-dimensional natural datasets. Second, online local learning reduced pairwise sample diversity even as it lowered average heat release, meaning practitioners must watch for mode narrowing. Third, physical analog circuits exhibit device mismatch and drift that can introduce exactly the residual offsets and temporal correlations Wang and Deng identify as the most damaging error modes.

Even so, the paper closes an important gap between physics and learning theory. By showing that a generative thermodynamic computer can learn through immediate local correlations and dissipate less heat as a result, Wang and Deng move thermodynamic AI one step closer to self-contained physical learning systems.