Research map / Probabilistic and scientific modeling / Dynamical systems
Neural Embeddings Rank: Aligning 3D latent dynamics with movements
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 Neural Embeddings Rank (NER), a contrastive-ranking dimensionality reduction method that embeds neural spiking dynamics into 3D latent space aligned with continuous hand movements, and evaluates it against six methods on monkey M1, PMd, and S1 reaching data.
Problem
Low-dimensional latent dynamics extracted from neural population activity often fail to align with continuous movement parameters (position, velocity, direction) and cannot be visualized compactly in 3D or decoded consistently across sessions and brain areas.
Approach
Modifies CEBRA's contrastive loss into a rank-based regression loss: each anchor's embeddings are ranked by continuous label distance (Rank-N-Contrast style), so all batch embeddings contribute as positives/negatives, addressing label imbalance and continuous-label structure without extra hyperparameters.
Key results
- NER explained 86% of velocity variance across ten M1 sessions versus 35% for the next best method (piVAE).
- Linear decoders trained on one session decoded velocity, position, and direction across dates and brain areas at 64%, 88%, and 90%, while CEBRA and piVAE gave negative velocity variance in cross-session M1 decoding.
- In S1, NER velocity decoding reached 86% vs 47% for the next best method.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NER instead of CEBRA or piVAE for linear velocity decoding | 3D neural embeddings · M1, same-session | explained variance: 0.71, 0.1 vs -3.08, 0.6; -1.21, 1.4 | Helped |
| Use NER instead of CEBRA or piVAE for linear velocity decoding | 3D neural embeddings · M1, cross-session | explained variance: 0.64, 0.1 vs -3.22, 1.0; -2.05, 1.6 | Helped |
| Use NER instead of CEBRA or piVAE for position decoding | 3D neural embeddings · M1, cross-session | explained variance: 0.88, 0.1 vs 0.71, 0.1; 0.23, 0.4 | Helped |
| Use NER instead of CEBRA or piVAE for direction decoding | 3D neural embeddings · M1, cross-session | accuracy: 90, 5.9 vs 79, 9.0; 43, 15.2 | Helped |
| Use NER instead of CEBRA or piVAE for linear velocity decoding | 3D neural embeddings · PMd, cross-session | explained variance: 0.74, 0.0 vs -3.10, 0.3; -2.78, 2.2 | Helped |
| Use NER instead of CEBRA or piVAE for linear velocity decoding | 3D neural embeddings · S1, same-session | explained variance: 0.67, 0.0 vs -4.59, 1.3; -0.38, 0.3 | Helped |
| Use NER instead of CEBRA or piVAE for position decoding | 3D neural embeddings · S1, cross-session | explained variance: 0.77, 0.1 vs 0.45, 0.1; 0.51, 0.1 | Helped |
| Use NER rather than CEBRA for curved and straight movement embeddings | M1 neural dynamics · Curved reaching task, combined straight-curve directions | explained variance: 81% vs 66% | Helped |
| Use NER rather than CEBRA for curved movement embeddings | M1 neural dynamics · Curved reaching task, combined curve-curve directions | explained variance: 81% vs 61% | Helped |
Surprising findings
- CEBRA and piVAE linear velocity decoders yielded negative explained variance in cross-session M1 decoding, while NER remained positive across all conditions.
- Both NER and CEBRA failed when all 108 MC_Maze movement conditions were trained simultaneously.
Limitations
- Both NER and CEBRA fail when many (108) movement conditions are trained together; 3D latent space may be insufficient for complex movements like handwriting or speech.
- Evaluated only on macaque reaching data with limited sessions and monkeys; no theoretical analysis.
Experimental setup
Datasets: Chewie center-out reaching M1/PMd (Dryad, Miller/Shenoy labs) Han center-out reaching S1 (Dryad) MC_Maze (DANDI 000128 via NLB)
Benchmarks: center-out reaching M1, PMd, S1 sessions MC_Maze curved reaching
Metrics: explained variance (R-squared) of velocity and position direction classification accuracy correlation between aligned averaged latent dynamics
Baselines: CEBRA piVAE PCA dPCA UMAP with labels UMAP without labels
Builds on: CEBRA Rank-N-Contrast piVAE UMAP PCA
Closest papers by meaning
- Meta-Dynamical State Space Models for Integrative Neural Data Analysis (ICLR 2025 · similarity 0.86)
- Exploring Behavior-Relevant and Disentangled Neural Dynamics with Generative Diffusion Models (NeurIPS 2024 · similarity 0.85)
- Latent Diffusion for Neural Spiking Data (NeurIPS 2024 · similarity 0.85)
- Neural-Behavioral Representation of Natural Whole-body Movement in Monkeys (NeurIPS 2026 · similarity 0.85)
- BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation (ICLR 2025 · similarity 0.85)
- Modeling Dynamic Neural Activity by combining Naturalistic Video Stimuli and Stimulus-independent Latent Factors (NeurIPS 2025 · similarity 0.85)
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.