Research map / Image generation / Fast diffusion sampling
Exploring Behavior-Relevant and Disentangled Neural Dynamics with Generative Diffusion Models
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 BeNeDiff, which learns a behavior-informed disentangled neural latent subspace from wide-field calcium imaging and uses a classifier-guided video diffusion model to generate behavior videos that reveal what each latent factor encodes.
Problem
Decoding models quantify how much behavioral information neural populations carry but do not reveal how behavior is encoded; neural populations across dorsal cortex regions show mixed selectivity, making factor-level interpretation of neural dynamics hard.
Approach
Combine a semi-supervised VAE with a total-correlation penalty (behavior-informed disentangled latent subspace) with a video diffusion model guided by a linear neural encoder and an NCE-style variance-contrast objective, so that activating one latent factor produces behavior videos specific to that factor.
Key results
- On VIS-Left, BeNeDiff's latent subspace reaches MIG 55.87% versus 43.98% for pi-VAE and 37.50% for SSL, at a cost in R^2 (75.41% vs 81.10% for SSL).
- On MOs-Left, MIG of 58.56% versus 42.20% for pi-VAE, while R^2 is 69.59% versus 76.65% for SSL.
- Single-factor decoding shows each of the six latent factors maps to a single behavior of interest (R^2 metric), in contrast to mixed selectivity of raw neural signals.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add behavior labels to the neural VAE (ablation) | BeNeDiff · VIS-Left and MOs-Left | MIG (%): 48.54 (± 0.23 ) -> 55.87 (± 0.26); 49.79 (± 0.23) -> 58.56 (± 0.29) | Helped |
| add the total-correlation penalty (ablation) | BeNeDiff · VIS-Left and MOs-Left | MIG (%): 38.13 (± 0.27) -> 55.87 (± 0.26); 40.74 (± 0.23) -> 58.56 (± 0.29) | Helped |
| add behavior labels and the total-correlation penalty instead of a standard VAE | BeNeDiff neural LVM · VIS-Right, Session-1 and Session-2 | MIG (%): 34.61 (± 0.30 ) -> 56.36 (± 0.29); 33.20 (± 0.27 ) -> 59.05 (± 0.26 ) | Helped |
| add behavior labels and the total-correlation penalty instead of a standard VAE | BeNeDiff neural LVM · VIS-Right, Session-1 and Session-2 | R² (%): 77.79 (± 0.20 ) -> 73.74 (± 0.24 ); 78.68 (± 0.21 ) -> 71.13 (± 0.29 ) | Hurt |
| add behavior labels and the total-correlation penalty instead of a standard VAE | BeNeDiff neural LVM · VIS-Right, Session-1 and Session-2 | RMSE: 48.94 (± 0.18 ) -> 55.17 (± 0.19 ); 49.54 (± 0.22 ) -> 54.27 (± 0.19 ) | Hurt |
| use BeNeDiff instead of SSL | neural LVM · VIS-Left, Session-1 | MIG (%): 37.50 (± 0.20 ) -> 55.87 (± 0.26) | Helped |
| use BeNeDiff instead of CEBRA | neural LVM · VIS-Left, Session-1 | MIG (%): 40.12 (± 0.24 ) -> 55.87 (± 0.26) | Helped |
| use BeNeDiff instead of pi-VAE | neural LVM · VIS-Left, Session-1 | MIG (%): 43.98 (± 0.29) -> 55.87 (± 0.26) | Helped |
| use BeNeDiff instead of SSL | neural LVM · MOs-Left, Session-1 | MIG (%): 36.89 (± 0.23 ) -> 58.56 (± 0.29) | Helped |
| use BeNeDiff instead of CEBRA | neural LVM · MOs-Left, Session-1 | MIG (%): 37.94 (± 0.23 ) -> 58.56 (± 0.29) | Helped |
| use BeNeDiff instead of pi-VAE | neural LVM · MOs-Left, Session-1 | MIG (%): 42.20 (± 0.28) -> 58.56 (± 0.29) | Helped |
| use BeNeDiff instead of SSL | neural LVM · VIS-Left and MOs-Left | R² (%): 81.10 (± 0.26 ) -> 75.41 (± 0.24); 76.65 (± 0.30 ) -> 69.59 (± 0.22) | Hurt |
Surprising findings
- Behavior-informed latent subspace trades some neural reconstruction R^2 for higher disentanglement, with reconstruction dropping relative to the self-supervised VAE in several regions.
Limitations
- Trade-off between disentanglement (MIG) and neural reconstruction quality requires per-region, per-session hyperparameter search.
- Neural encoder used as guidance classifier is linear, which may limit guidance performance.
- Validated on a single widefield calcium imaging dataset from one mouse task setting with two sessions.
Experimental setup
Datasets: Widefield calcium imaging dataset (head-fixed mouse visual decision-making, two sessions)
Benchmarks: Widefield calcium imaging dataset (two sessions)
Metrics: R-squared (R^2 %) RMSE MIG cross-validation R^2 (cvR^2)
Baselines: SSL CEBRA pi-VAE Naive latent manipulation Classifier-free guidance Standard VAE
Builds on: variational autoencoder classifier guidance noise-contrastive estimation classifier-free guidance MIG
Closest papers by meaning
- BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation (ICLR 2025 · similarity 0.88)
- Time-Evolving Dynamical System for Learning Latent Representations of Mouse Visual Neural Activity (NeurIPS 2025 · similarity 0.87)
- Modeling Dynamic Neural Activity by combining Naturalistic Video Stimuli and Stimulus-independent Latent Factors (NeurIPS 2025 · similarity 0.86)
- Modeling Latent Neural Dynamics with Gaussian Process Switching Linear Dynamical Systems (NeurIPS 2024 · similarity 0.86)
- Latent Diffusion for Neural Spiking Data (NeurIPS 2024 · similarity 0.86)
- Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series (ICML 2025 · similarity 0.86)
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.