Research map / Probabilistic and scientific modeling / Dynamical systems

Identifying Neural Dynamics Using Interventional State Space Models

ICML 2025 · Probabilistic and scientific modeling › Dynamical systems · approach: Latent models of neural activity

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes interventional state-space models (iSSM) that decouple perturbed latents from their upstream dependencies, with identifiability theory and experiments on simulated motor dynamics and real calcium and macaque micro-stimulation data.

Problem

Standard state space models fit to neural data capture only observational statistical associations and cannot predict neural dynamics under causal perturbations, and their parameters are often non-identifiable.

Approach

Model interventions as changing the dynamics: when a latent is perturbed, its incoming edges from upstream latents are cut (column of the dynamics matrix ignored), with a sparse Laplace prior on the intervention matrix B; combined with identifiability theory based on interventional independence, enabling generalization to unseen interventions.

Key results

• On simulated motor cortex data, iSSM recovers true latents (reconstruction correlation improves with more interventions) while SSM does not, especially in high-dimensional observations.
• On mouse ALM photostimulation data, iSSM reconstructs training and test trials better than SSM across latent dimensions.
• On macaque dlPFC micro-stimulation data, iSSM outperforms SSM in train and test reconstruction accuracy and yields more consistent B across random initializations.

Measured effects reported in the paper

Limitations

• Generative model restricted to linear latent dynamics with nonlinear emissions; nonlinear dynamics lack theory.
• Assumes interventions act directly on latents that are causally linked to them, which needs experimental validation.
• Relies on approximate variational inference with LSTM recognition network, which may amplify approximation gap.

Experimental setup

Datasets: simulated rotational dynamics and dynamic attractor data mouse ALM calcium imaging with photostimulation (Daie et al., 2021) macaque dlPFC electrophysiology with micro-stimulation (Nejatbakhsh et al., 2023)

Benchmarks: simulated motor cortex models mouse ALM photostimulation dataset macaque dlPFC micro-stimulation dataset

Metrics: reconstruction accuracy (correlation) consistency score of B across initializations

Baselines: SSM (observational state space model)

Builds on: linear dynamical systems causal representation learning Ahuja et al. identifiability Lippe et al. interventional identifiability Hyttinen et al. unordered pairs condition

Closest papers by meaning

• Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series (ICML 2025 · similarity 0.89)
• Meta-Dynamical State Space Models for Integrative Neural Data Analysis (ICLR 2025 · similarity 0.88)
• A scalable generative model for dynamical system reconstruction from neuroimaging data (NeurIPS 2024 · similarity 0.87)
• Probabilistic Decomposed Linear Dynamical Systems for Robust Discovery of Latent Neural Dynamics (NeurIPS 2024 · similarity 0.87)
• Inference of Neural Dynamics Using Switching Recurrent Neural Networks (NeurIPS 2024 · similarity 0.86)
• Tangent Space Causal Inference: Leveraging Vector Fields for Causal Discovery in Dynamical Systems (NeurIPS 2024 · similarity 0.86)