Inferring brain plasticity rule under long-term stimulation with structured recurrent dynamics
Zhichao Liang, Jingzhe Lin, Xinyi Li, Guanyi Zhao, Quanying Liu
TL;DR
This paper introduces STEER, a novel framework that models long-term brain plasticity as a latent dynamical law, enabling inference of plasticity rules and improving intervention strategies through data-driven analysis.
Contribution
The paper presents STEER, a dual-timescale recurrent model that disentangles neural activity from plasticity, allowing for interpretable inference of brain plasticity rules from long-term stimulation data.
Findings
STEER accurately recovers interpretable plasticity equations.
It predicts network adaptation under unseen stimulation schedules.
Supports design of improved brain stimulation protocols.
Abstract
Understanding how long-term stimulation reshapes neural circuits requires uncovering the rules of brain plasticity. While short-term synaptic modifications have been extensively characterized, the principles that drive circuit-level reorganization across hours to weeks remain unknown. Here, we formalize these principles as a latent dynamical law that governs how recurrent connectivity evolves under repeated interventions. To capture this law, we introduce the Stimulus-Evoked Evolution Recurrent dynamics (STEER) framework, a dual-timescale model that disentangles fast neural activity from slow plastic changes. STEER represents plasticity as low-dimensional latent coefficients evolving under a learnable recurrence, enabling testable inference of plasticity rules rather than absorbing them into black-box parameters. We validate STEER with four benchmarks: synthetic Lorenz systems with…
Peer Reviews
Decision·ICLR 2026 Poster
* The framework seems principled in its design and the approach overcomes disadvantages of certain prior meta-learning approaches by not assuming a specific functional form of the underlying learning rule (e.g., not restricted to just variants of Oja's rule). * The presentation and figures are mostly clear. * The experimental evaluations span both synthetic tasks and real data, which is important for such works.
* The use of DSA is good but you only report a single DSA value of 0.63 with no baseline or control/chance value. DSA is a relative metric, as emphasised in the original paper. Thus, it is not possible to know whether a DSA of 0.63 is good without a baseline comparison. Could the authors provide a baseline on shuffled values of the implicit factor, for example, and show that the actual inferred dynamics have a higher DSA score with the true dynamics? * In Fig. 4f I am not sure you can claim that
1. **Motivation:** The paper addresses an underexplored area by extending short-term plasticity modeling toward the longer timescales of circuit reorganization. The proposed framework offers a structured, data-driven way to describe how stimulation may gradually reshape network connectivity. 2. **References:** The related research is carefully reviewed, linking established neuroscience findings on Hebbian and homeostatic mechanisms with recent machine learning approaches for recurrent dynamics a
1. **Evaluation:** Results on the BCM and Parkinson’s DBS datasets appear modest, and there is a visible mismatch between the ground truth in Fig. 4(a) and the model output in Fig. 4(b), suggesting partial recovery of the connection change. 2. **Baselines:** Only one baseline (MD-SSM) is evaluated on the BCM simulation. Including other relevant machine learning approaches discussed in the related work would strengthen the empirical comparison. 3. **Benchmark:** The evaluation spans two synthe
1. The paper treats long-horizon plasticity as a latent dynamical law, rather than unstructured parameter drift. 2. The model separates fast within-session dynamics and slow across-session evolution.
1. The dynamical systems have input weights and readouts, which can also encode some information of connectivity. Therefore, it is still unsure if the connectivity recovered by the model is true or believable. (This may have been claimed by the author in the limitation part, but it remains a substantive concern.) 2. The authors stated that the proposed method enforces an identifiable separation between fast within-session responses and slow network reconfiguration. But there’s no theorem or abla
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsNeural dynamics and brain function · Neurological disorders and treatments · Functional Brain Connectivity Studies