Research map / Probabilistic and scientific modeling / Dynamical systems
Identifying interactions across brain areas while accounting for individual-neuron dynamics with a Transformer-based variational autoencoder
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 GLM-Transformer, a hybrid model pairing a Transformer-based VAE for trial-varying individual-neuron dynamics with a low-rank point-process GLM coupling term to identify directed cross-area interactions in spike trains.
Problem
Identifying directed interactions between neural populations from spike trains is confounded by trial-to-trial variability in background dynamics (e.g., locomotion, arousal) that is shared across areas and can be misattributed to coupling.
Approach
Combine a Transformer VAE that infers a per-trial latent and decodes smooth B-spline individual-neuron dynamics with a low-rank, raised-cosine-filtered GLM coupling term, trained in a four-stage procedure so that background dynamics absorb shared fluctuations rather than coupling estimates.
Key results
- On the Allen dataset, the full model reaches test NLL 4.218e7 versus 4.264e7 without the trial-wise latent and 4.263e7 without the coupling term (ablation).
- In robustness simulations with no true coupling, GLM-Transformer keeps spurious coupling variance low while RRR, mDLAG, and single-neuron GLMs attribute shared gain fluctuations to coupling.
- Identifies strong feedforward coupling from V1 to LM and V1 to AL, consistent with the known visual hierarchy.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace the Transformer encoder with a bidirectional RNN (ablation) | GLM-Transformer · GLM Simulation | test negative log-likelihood: 3.008 × 10^7 -> 3.032 × 10^7 | Hurt |
| replace the Transformer encoder with a bidirectional RNN (ablation) | GLM-Transformer · EIF Simulation | test negative log-likelihood: 1.513 × 10^7 -> 1.515 × 10^7 | No clear effect |
| replace the Transformer encoder with a bidirectional RNN (ablation) | GLM-Transformer · Allen Visual Coding | test negative log-likelihood: 4.218 × 10^7 -> 4.225 × 10^7 | No clear effect |
| remove the trial-wise latent variable (ablation) | GLM-Transformer · GLM Simulation | test negative log-likelihood: 3.008 × 10^7 -> 3.035 × 10^7 | Hurt |
| remove the trial-wise latent variable (ablation) | GLM-Transformer · EIF Simulation | test negative log-likelihood: 1.513 × 10^7 -> 1.521 × 10^7 | Hurt |
| remove the trial-wise latent variable (ablation) | GLM-Transformer · Allen Visual Coding | test negative log-likelihood: 4.218 × 10^7 -> 4.264 × 10^7 | Hurt |
| remove the coupling term (ablation) | GLM-Transformer · GLM Simulation | test negative log-likelihood: 3.008 × 10^7 -> 3.065 × 10^7 | Hurt |
| remove the coupling term (ablation) | GLM-Transformer · EIF Simulation | test negative log-likelihood: 1.513 × 10^7 -> 1.537 × 10^7 | Hurt |
| remove the coupling term (ablation) | GLM-Transformer · Allen Visual Coding | test negative log-likelihood: 4.218 × 10^7 -> 4.263 × 10^7 | Hurt |
| remove the post-spike history filter (ablation) | GLM-Transformer · GLM Simulation | test negative log-likelihood: 3.008 × 10^7 -> 3.109 × 10^7 | Hurt |
| remove the post-spike history filter (ablation) | GLM-Transformer · EIF Simulation | test negative log-likelihood: 1.513 × 10^7 -> 1.603 × 10^7 | Hurt |
| remove the post-spike history filter (ablation) | GLM-Transformer · Allen Visual Coding | test negative log-likelihood: 4.218 × 10^7 -> 4.294 × 10^7 | Hurt |
Surprising findings
- A bidirectional RNN encoder performs comparably to the Transformer on the EIF and Allen datasets, suggesting the architecture choice matters little there.
Limitations
- Requires over a thousand trials to converge, about an order of magnitude more data than standard GLMs.
- Assumes fixed trial length, limiting applicability to variable-length data.
- No theoretical guarantees on training performance and no statistical inference (uncertainty) for coupling estimates.
Experimental setup
Datasets: Allen Institute Visual Coding dataset synthetic GLM spike trains synthetic EIF simulations
Benchmarks: Allen Institute Visual Coding dataset synthetic GLM simulation EIF simulation
Metrics: test negative log-likelihood recovery of ground-truth coupling percentage of variance attributed to coupling
Baselines: reduced-rank regression mDLAG single-neuron GLM single-neuron GLM with trial-wise gain GLM-Transformer without trial-varying dynamics bidirectional RNN encoder variant
Builds on: Deep Random Splines Neural Data Transformer Generalized linear model point process Gaussian Process Factor Analysis reduced-rank regression
Closest papers by meaning
- Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data (ICLR 2024 · similarity 0.89)
- Multi-modal Gaussian Process Variational Autoencoders for Neural and Behavioral Data (ICLR 2024 · similarity 0.88)
- Accurate Identification of Communication Between Multiple Interacting Neural Populations (ICML 2025 · similarity 0.88)
- BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation (ICLR 2025 · similarity 0.88)
- One-hot Generalized Linear Model for Switching Brain State Discovery (ICLR 2024 · similarity 0.87)
- CREIMBO: Cross-Regional Ensemble Interactions in Multi-view Brain Observations (ICLR 2025 · 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.