Artemy Kolchinsky

How do time series foundation models forecast unseen dynamical systems? In new experiments, we find that small transformers learn to approximate transfer operators in-context. (1/N) arxiv.org/abs/2602.18679

@RomeoLupascu @pfau @m_aguilera_ @ito_sosuke To be precise, the term "neural data" refers to actual brain recordings (spike train dataset). In our paper, we did not apply our method to neural networks

Well, when you put "spin glass" and "neural net" in the same sentence that says something about what is our state of understanding of neural nets doesn't it? More precise, the state of understanding a trained neural net is close to zero. What can go wrong when we build systems that we don't understand or be able to predict their behavior then let them loose in the world while hyping that they are our next gods. Have we just went insane? Just thinking...

Our paper w/@m_aguilera_& @ito_sosuke is out in PRL. We combine ideas from stochastic thermodynamics & nonequilibrium Maximum Entropy to quantify forces + dissipation in high-dimensional nonequilibrium systems, including spin glasses and neural data journals.aps.org/prl/abstract/1…

Generative thermodynamic computing Diffusion models are powerful generative tools, but they come with a hidden cost: every denoising step requires a digital neural network, artificially injected noise, and substantial energy consumption. Yet physics offers an alternative—what if the noise needed for generation arose naturally from thermal fluctuations, and the denoising process was physically enacted rather than simulated? Stephen Whitelam introduces exactly this: a generative modeling framework for thermodynamic computing. Instead of using neural networks to transform noise into structure, the approach encodes denoising information directly in the energy landscape of a physical system evolving under Langevin dynamics. The training principle is elegant: observe noising trajectories (structured data degrading into noise), then adjust the system's couplings via gradient descent to maximize the probability that a thermodynamic computer would generate the reverse—structure from noise. This process has a beautiful physical interpretation: it minimizes the heat emission and entropy production of the generative process. In a proof-of-concept simulation with 784 visible units and 512 hidden units trained on just three MNIST digits, the thermodynamic computer learns to transform noise into recognizable digit-like structures through physical dynamics alone—no external control or pseudorandom numbers required. The energy implications are striking: the simulated thermodynamic computer emits ~2,900 kᵦT of heat per generation, compared to ~5 × 10¹⁴ kᵦT for a digital neural network doing equivalent denoising—a difference of more than 10 orders of magnitude. The message is compelling: by grounding generative modeling in thermodynamic principles, we can design systems where computation emerges from physics itself, opening paths toward autonomous, energy-efficient generation that could fundamentally change how we think about the hardware of machine learning. Paper: journals.aps.org/prl/abstract/1…

The chemosensing accuracy of 𝘌. 𝘤𝘰𝘭𝘪 cells is shown to be limited by internal noise in signal processing, rather than the stochasticity of molecule arrivals at their receptors, contrary to long-held understanding in the field. nature.com/articles/s4156…

Our paper was published on the relationship between thermodynamic driving and eigenvalues in Markovian master equations. It proves a weaker version of a beautiful conjecture proposed by Uhl and Seifert. Led by Guo-Hua Xu , with Jean-Charles Delvenne and @ito_sosuke

ポスドクのGuohuaさんの論文がPRLのEditors' suggestionに選ばれました。ArtemyさんとJean-Charlesさんとの共著です。マスター方程式の遷移レートの固有値に関して、熱力学的な制約から楕円の中にいる必要があるという定理を導出しています。 doi.org/10.1103/z4t2-1…

We are hiring 1-2 postdocs for next year! Please reach out if interested in: 1. Inter-brain information translation & control 2. Characterizing brain states via thermodynamics & control theory 3. Comparing qualia structures across individuals & identifying neural substrates See the new lab website for the details! oizumi-lab-website.vercel.app/join/postdocs

In Fall 01968, @stewartbrand put the Whole Earth Catalog out into the world. 55 years later, our friends at @grayareaorg and the @internetarchive are launching the Whole Earth Index: a nearly complete free online archive of Whole Earth publications. wholeearth.info

How do nonlinear oscillations generate thermodynamic dissipation? Our new arXiv preprint with @ito_sosuke and @oizumim introduces a Koopman-based framework that breaks down dissipation in nonlinear dynamics into contributions from oscillatory modes. arxiv.org/abs/2510.21340

A new computational algorithm makes it possible to quantify the transfer entropy for any stochastic model without the need for approximations go.aps.org/47ooZGG

Submissions open for a really interesting workshop on quantum and stochastic thermodynamics - Kyoto, Japan, Dec 8-12 indico.yukawa.kyoto-u.ac.jp/event/68/

【発表申込締切：8/31】12/8〜12/12に京都大学基礎物理学研究所で開催される国際研究会 "Kyoto Workshop on Quantum Thermodynamics and Stochastic Thermodynamics 2025" の発表申込締切が近づいてまいりました。皆様、ぜひ奮ってご参加ください。indico.yukawa.kyoto-u.ac.jp/event/68/

The work was done by Stuart Bartlett, @Ghoshal_G, @AdamFrank4, and @artemyte at @SETIInstitute, @YorkUniversity, @AucklandUni, @FloridaTech, Universitat Pompeu Fabra, and @UofR.

A new Roadmap discusses how information influences abiogenesis and living systems. It explores how informational constraints may help gauge planet habitability — and how informational #biosignatures could indicate life beyond Earth. 🔗 go.aps.org/4mJdjD0 #Astrobiology

Postdoc opportunity! Join us in heavenly Vancouver (Canada) to develop fundamental nonequilibrium stat mech, thermo, and info theory applied to biomolecular machines and in close collaboration with experiment. Details: sfu.ca/physics/sivakg…

🚨Job alert!🚨 Come join us at the Inverse Complexity Lab! @invcomplexity We’re hiring a post-doctoral researcher to join our group at IT:U, Linz, Austria. skewed.de/lab/call.html Deadline is 30 Nov 2024. 1/8

Graduating PhDs and postdocs in AI, ML, Cogsci, or related areas: apply to work with me and others on AI models of visual and multimodal reasoning. Two years of funding with possible extension to a third year. Application deadline November 22. santafe.edu/about/jobs/pos…

We are hiring! Join our Department of Chemistry at @UMassBoston as a tenure-track Assistant Professor (Inorganic Materials). Please spread the word. bit.ly/4eBeD6z