DurstewitzLab

@Esychology @wgilpin0 Why it's so good in capturing the long-term behavior zero-shot we haven't fully understood so far, but a major part of it is likely the control-theoretic training techniques we are using, plus perhaps the ALRNNs as experts (see ablation studies in sec. 4.5 of paper)

@DurstewitzLab @wgilpin0 Could you explain how this can predict the long term dynamics of chaotic systems like the ones given by Lorenz equations ?

Tomorrow Christoph will present DynaMix, the first foundation model for dynamical systems reconstruction, at #NeurIPS2025 Exhibit Hall C,D,E #2303

RT @sciqst: @DurstewitzLab @GeorgiaKoppe @BrunoAverbeck Captivating perspective! The flexibility and adaptability of the human brain indeed…

RT @emrahdma: Now that the AI overlords all started publicly admitting defeat with the current architecture, I believe we are going to see…

Unlike current AI systems, brains can quickly & flexibly adapt to changing environments. This is the topic of our perspective in Nature MI (rdcu.be/eSeif), where we relate dynamical & plasticity mechanisms in the brain to in-context & continual learning in AI. #NeuroAI

Despite being extremely lightweight (only 0.1% of params, 0.6% training corpus size, of closest competitor), it also outperforms major TS foundation models like Chronos variants on real-world TS short-term forecasting with minimal inference times (0.2%) ...

Our #DynamicalSystems #FoundationModel was accepted to #NeurIPS2025 with outstanding reviews (6555) – first model which can *0-shot*, w/o any fine-tuning, forecast the *long-term statistics* of time series provided a context. Test it on #HuggingFace: huggingface.co/spaces/Durstew… ...

Relevant publications: nature.com/articles/s4158… openreview.net/pdf?id=Vp2OAxM… proceedings.mlr.press/v235/brenner24… nature.com/articles/s4146…

We have openings for several fully-funded positions (PhD & PostDoc) at the intersection of AI/ML, dynamical systems, and neuroscience within a BMFTR-funded Neuro-AI consortium, at Heidelberg University & Central Institute of Mental Health (see below): einzigartigwir.de/en/job-offers/…

Our new preprint compares naïve baselines, network models (incl. PLRNN-based SSMs), and Transformers on 3x40‑day EMA+EMI datasets. PLRNNs gave the most accurate forecasts, yielded interpretable networks, and flagged “sad” & “down” as top leverage points. doi.org/10.1101/2025.0…

Got prov. approval for 2 major grants in Neuro-AI & Dynamical Systems Recons., on learning & inference in non-stationary environments, OOD generalization, and DS foundation models. To all AI/math enthusiasts: Expect job announcements (PhD/PostDoc) soon! Feel free to get in touch.

Mathematics & kindness

We wrote a little #NeuroAI piece about in-context learning & neural dynamics vs. continual learning & plasticity, both mechanisms to flexibly adapt to changing environments: arxiv.org/abs/2507.02103 We relate this to non-stationary rule learning w rapid jumps. Feedback welcome!

Fantastic work by Florian Bähner, Hazem Toutounji, Tzvetan Popovand & many others - I'm just the person advertising!

How do animals learn new rules? By systematically testing diff. behavioral strategies, guided by selective attn. to rule-relevant cues: rdcu.be/etlRV Akin to in-context learning in AI, strategy selection depends on the animals' "training set" (prior experience).

RT @russo_eleon: Into population dynamics? Coming to #CNS2025 but not quite ready to head home? Come join us! at the Symposium on "Neural…

@pastramimachine So far we have released the code only for reviewing purposes, but will soon make it public ...

Can time series #FoundationModels like Chronos zero-shot generalize to unseen #DynamicalSystems (DS)? No, they cannot. But *DynaMix* can, the first FM based on principles of DS reconstruction, capturing the long-term evolution of out-of-domain DS: arxiv.org/pdf/2505.13192… (1/6)

We dive a bit into the reasons why current time series FMs not trained for DS reconstruction fail, and conclude that a DS perspective on time series forecasting & models may help to advance the #TimeSeriesAnalysis field. (6/6)

Remarkably, DynaMix not only generalizes zero-shot to novel DS, but it can even generalize to new initial conditions and regions of state space not covered by the in-context information. (5/6)