TL;DR
This paper introduces a meta-learning approach to infer shared neural dynamics across different recordings and tasks, enabling rapid adaptation and analysis of neural data from multiple sources.
Contribution
It proposes a novel meta-dynamical state space model that captures variability across neural recordings on a low-dimensional manifold for improved latent dynamics inference.
Findings
Effective in few-shot reconstruction of synthetic systems
Successfully applied to neural recordings during arm reaching tasks
Facilitates rapid learning of neural dynamics from new data
Abstract
Learning shared structure across environments facilitates rapid learning and adaptive behavior in neural systems. This has been widely demonstrated and applied in machine learning to train models that are capable of generalizing to novel settings. However, there has been limited work exploiting the shared structure in neural activity during similar tasks for learning latent dynamics from neural recordings. Existing approaches are designed to infer dynamics from a single dataset and cannot be readily adapted to account for statistical heterogeneities across recordings. In this work, we hypothesize that similar tasks admit a corresponding family of related solutions and propose a novel approach for meta-learning this solution space from task-related neural activity of trained animals. Specifically, we capture the variabilities across recordings on a low-dimensional manifold which…
Peer Reviews
Decision·ICLR 2025 Spotlight
* **Integration of Multi-Session Neural Recordings:** The proposed framework effectively addresses the challenges of integrating (potentially heterogeneous) neural recordings. It achieves this through a parameterization of latent dynamics using a low-dimensional dynamical embedding to capture variations across datasets. While they are not the first report to do so, it is an important direction/application. * **Learning a Shared Dynamical Structure:** The model learns a shared dynamical struc
**Missing Benchmarks & Issues in Related Works** - There are some inaccurate statements in the introduction; for example, "CEBRA (Schneider et al., 2023) and CS-VAE (Yi et al., 2022) extract latent representations but do not learn underlying dynamics." Schneider et al. specifically learn the underlying dynamics rather than a priori prescribing them, as other algorithms do (e.g., SIDS, LFADS). CS-VAE assumes the same shared underlying dynamics and uses witching linear dynamical systems (SLDS) t
- The paper is well-motivated. The introduction frames the paper well by highlighting surrounding literature and results on shared structure across subjects/tasks in task-trained models. Then section 2 provides a compelling example of the limitations of a shared latent dynamics model, with dataset-specificity only in the emission likelihood. - The paper is generally well-written. - The synthetic results are strong and form a cohesive story with Section 2 - The motor cortex recording analysis i
While the authors do explore ablations and specific perspectives on their own model class, a comparison to other models is missing. There are two fronts to this comparison: - **Theoretical and modeling connections**, investigating how their SSM relates to known models. Can you cast known models as special cases of yours? For instance the shared dynamics motifs (Driscoll et al, 2024), how would they best fit within this framework? Alternatively, the embedding analysis of (Cotler et al., 2023), ho
The paper addresses a long-standing problem in the field of neural data analysis, and one that will only become more pronounced as data collection methods improve. The approach is novel and well-motivated, and provides a tool for not only nicely reconstructing neural dynamics but also providing interpretable model components that will be crucial for scientific understanding. Additionally, the paper is well-written. The proof of concept presented in figs 2+3 is extremely helpful for introducting
While the authors generally do a great job keeping track of all the notation, I find myself confused about how different trials are handled. It seems that trial averages over, say, certain reach directions are being modeled, rather than individual trials (perhaps I missed this somewhere). But how, then, are the different conditions handled in the notation y_{1:T}^{1:M}? I became confused about this in section 3.2 regarding the read-in network and the dynamical embeddings. Is there a single dynam
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
