Research map / Label-efficient and robust learning / Knowledge distillation
BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation
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 BLEND, a model-agnostic privileged knowledge distillation framework that trains a teacher on neural activity plus behavior and distills a neural-only student, improving NDT and LFADS on NLB'21 and NeuPRINT on transcriptomic identity prediction.
Problem
Neural population dynamics models usually cannot use behavior at inference, and paired neural-behavior data is often absent; the question is how to exploit behavior during training while relying only on neural activity at test time.
Approach
Treat behavior as privileged information: a teacher takes masked neural activity plus behavior, and a neural-only student is trained with masked reconstruction plus a distillation loss (hard, soft, feature, or correlation), making the approach agnostic to the base architecture.
Key results
- NDT-Distill-Best reaches Vel-R2 0.891 on MC-Maze (+14.4% over NDT) and 0.788 on Area2-Bump (+51.8%).
- LFADS-Distill-Best achieves PSTH-R2 0.615 on Area2-Bump (+44.7% over LFADS).
- NeuPRINT-Distill-Best reaches 0.571 subclass accuracy with multiple animals (+15.4%) and 0.588 in single animal (+15.7%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply behavior-guided correlation distillation to NDT | NDT · MC-Maze | Co-bps: 0.275 -> 0.310 (+12.7\%) | Helped |
| Apply behavior-guided correlation distillation to NDT (ablation) | NDT · MC-Maze | Vel-R²: 0.779 -> 0.891 (+14.4\%) | Helped |
| Apply behavior-guided soft distillation to NDT (ablation) | NDT · MC-Maze | Co-bps: 0.275 -> 0.259 (-5.8\%) | Hurt |
| Apply behavior-guided soft distillation to NDT (ablation) | NDT · MC-Maze | Vel-R²: 0.779 -> 0.881 (+13.1\%) | Helped |
| Apply behavior-guided soft distillation to LFADS (ablation) | LFADS · MC-Maze | Co-bps: 0.315 -> 0.282 (-10.5\%) | Hurt |
| Apply behavior-guided soft distillation to LFADS (ablation) | LFADS · MC-RTT | Vel-R²: 0.416 -> 0.443 (+6.5\%) | Helped |
| Apply behavior-guided hard distillation to LFADS (ablation) | LFADS · Area2-Bump | Vel-R²: 0.649 -> 0.837 (+29.0\%) | Helped |
| Apply behavior-guided hard distillation to NDT (ablation) | NDT · Area2-Bump | PSTH-R²: 0.290 -> 0.363 (+25.2\%) | Helped |
| Apply behavior-guided feature distillation to NDT (ablation) | NDT · Area2-Bump | PSTH-R²: 0.290 -> 0.481 (+65.9\%) | Helped |
| Apply behavior-guided correlation distillation to NeuPRINT (ablation) | NeuPRINT · Multiple Animals | Top-1 accuracy, EI: 0.748 -> 0.792 (+5.9\%) | Helped |
| Apply behavior-guided hard distillation to NeuPRINT (ablation) | NeuPRINT · Multiple Animals | Top-1 accuracy, subclass: 0.495 -> 0.571 (+15.4\%) | Helped |
| Apply behavior-guided soft distillation to NeuPRINT (ablation) | NeuPRINT · Single Animal | Top-1 accuracy, subclass: 0.508 -> 0.588 (+15.7\%) | Helped |
Surprising findings
- Soft distillation degraded NDT and LFADS on several metrics (e.g., LFADS-Soft-Distill Co-bps -10.5% on MC-Maze), showing distillation strategy effects are not uniform.
- Distillation gains for LFADS on MC-Maze and MC-RTT were small compared with NDT, which the authors attribute to LFADS capacity limits.
Limitations
- Behavior is concatenated along the feature dimension, a simple fusion that may not capture complex neural-behavior interactions.
- Assumes behavior is temporally continuous and aligned with neural activity at each time step.
- Distillation gains vary by base model and strategy, so strategy selection is task-specific.
Experimental setup
Datasets: MC-Maze MC-RTT Area2-Bump multi-modal calcium imaging dataset (V1, 4 mice)
Benchmarks: NLB'21 (MC-Maze, MC-RTT, Area2-Bump) multi-modal calcium imaging transcriptomic dataset
Metrics: Co-bps Vel-R2 PSTH-R2 top-1 accuracy
Baselines: LRNN LFADS NDT STNDT MINT pi-VAE PSID TNDM PCA UMAP LOLCAT NeuPRINT
Builds on: learning under privileged information knowledge distillation NDT LFADS NeuPRINT
Closest papers by meaning
- BRAID: Input-driven Nonlinear Dynamical Modeling of Neural-Behavioral Data (ICLR 2025 · similarity 0.89)
- Identifying interactions across brain areas while accounting for individual-neuron dynamics with a Transformer-based variational autoencoder (NeurIPS 2025 · similarity 0.88)
- POCO: Scalable Neural Forecasting through Population Conditioning (NeurIPS 2025 · similarity 0.88)
- Exploring Behavior-Relevant and Disentangled Neural Dynamics with Generative Diffusion Models (NeurIPS 2024 · similarity 0.88)
- Towards a "Universal Translator" for Neural Dynamics at Single-Cell, Single-Spike Resolution (NeurIPS 2024 · similarity 0.87)
- Cross-Modal Representational Knowledge Distillation for Enhanced Spike-informed LFP Modeling (NeurIPS 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.