Research map / Probabilistic and scientific modeling / Dynamical systems
Partial observation can induce mechanistic mismatches in data-constrained models of neural dynamics
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.
Shows theoretically and numerically that fitting low-dimensional data-constrained RNNs/LDS models to partial observations of teacher networks with non-normal or low-rank connectivity can produce spurious line attractors and limit cycles, while normal/symmetric line attractors are recovered.
Problem
Whether data-constrained RNN surrogate models fit to partially observed neural recordings recover the correct mechanistic (attractor) structure of the true circuit, given that most neurons in a circuit are unobserved.
Approach
Teacher-student analysis showing that subsampling neurons biases low-dimensional students toward large eigenvalues, so non-normal teachers (functionally feedforward chains, null-overlap low-rank networks) yield spurious line/hyperplane attractors and limit cycles, while symmetric line attractors are faithfully recovered; the MAP estimate has closed form in the infinite-time limit.
Key results
- Symmetric teacher line attractors are recovered by the student with eigenvalues bounded by teacher spectrum (lambda_hat_1 >= lambda_1 - O(eps D/d)).
- For a feedforward chain teacher with D to infinity at fixed d, the student dynamics approach a line attractor with leading eigenvalue approaching 1.
- Rank-2 and rank-3 null-overlap teachers with 5% observation yield spurious line and plane attractors; at 10% subsampling MAP and CORNN/BPTT/FORCE fits produce spurious limit cycles.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fit a low-dimensional LDS to recordings from only 5% of neurons | Line-attractor and functionally feedforward integrator networks · Sensory integration task | mechanistic recovery | Mixed |
| Observe only a subset of neurons and fit a smaller student to the feedforward-chain teacher | MAP student network · Noise-driven feedforward chain | attractor structure | Hurt |
| Observe only a subset of neurons from a non-normal rank-2 teacher | MAP student network · Noise-driven low-rank teacher, rank 2 | slow directions | Hurt |
| Observe only a subset of neurons from a non-normal rank-3 teacher | MAP student network · Noise-driven low-rank teacher, rank 3 | slow directions | Hurt |
| Partially observe a symmetric teacher network | MAP student network · Noise-driven normal dynamics | timescale recovery | No clear effect |
| Partially observe a nonlinear low-rank non-normal teacher | MAP student network · Nonlinear noise-driven low-rank teacher | spurious limit-cycle discovery | Hurt |
| Use CORNN, BPTT, or FORCE instead of MAP to infer student dynamics | Student networks fit to nonlinear low-rank teachers · Nonlinear noise-driven low-rank teacher | attractor structure | Hurt |
Surprising findings
- A motivating LDS fit to a functionally feedforward chain reaches a line attractor score above 6, indistinguishable from a true line attractor in both data fit and flow fields.
- Explained-variance-based latent dimensionality selection biases toward attractor-like solutions.
Limitations
- Analysis restricted to single-trial settings with known or white-noise inputs; unobserved inputs not addressed.
- Analytical results mostly for linear networks, infinite observation time, and specific teacher structures (feedforward chains, null-overlap low-rank).
Experimental setup
Datasets: simulated teacher RNN activity
Metrics: eigenvalue spectrum line attractor score time constants flow field agreement
Baselines: FORCE BPTT CORNN LDS variational inference
Builds on: latent linear dynamical system fitting (variational inference) CORNN FORCE backpropagation through time
Closest papers by meaning
- Inference of Neural Dynamics Using Switching Recurrent Neural Networks (NeurIPS 2024 · similarity 0.88)
- Mechanistic Interpretability of RNNs emulating Hidden Markov Models (NeurIPS 2025 · similarity 0.88)
- Dynamical phases of short-term memory mechanisms in RNNs (ICML 2025 · similarity 0.88)
- Feedback control guides credit assignment in recurrent neural networks (NeurIPS 2024 · similarity 0.87)
- BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation (ICLR 2025 · similarity 0.87)
- The Bayesian sampling in a canonical recurrent circuit with a diversity of inhibitory interneurons (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.