Research map / Vision-language and multimodal / CLIP and zero-shot
Neural Representational Consistency Emerges from Probabilistic Neural-Behavioral Representation Alignment
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 PNBA, a probabilistic VAE-based neural-behavioral representation alignment framework that yields zero-shot cross-subject preserved neural representations in monkey M1/PMd and mouse V1.
Problem
Explaining how neural circuits produce consistent functional representations across individuals despite hierarchical trial, session, and subject variability, with zero-shot validation rather than manual post-hoc alignment.
Approach
Combines a probabilistic distribution-level matching objective between neural and behavioral encoders with bidirectional generative (VAE/ELBO) reconstruction constraints to prevent representation collapse, using one shared cross-subject network that standardizes variable neuron counts via adaptive pooling and interpolation.
Key results
- On held-out new subjects, PNBA reaches R = 0.9302 in M1 versus 0.7060 for MEME and 0.3627 for FA+amLDS.
- In mouse V1, PNBA achieves R = 0.8705 on new subjects versus 0.5980 for MEME.
- Zero-shot running-speed decoding from V1 reaches R^2 = 0.888 with a GRU decoder.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace MEME with PNBA | Cross-subject alignment on M1 training subjects · M1 training subjects | Pearson correlation (R): 0.7756 -> 0.9465 | Helped |
| Replace MEME with PNBA | Cross-subject alignment on M1 new subjects · M1 new subjects | Pearson correlation (R): 0.7060 -> 0.9302 | Helped |
| Replace MEME with PNBA | Cross-subject alignment on PMd training subjects · PMd training subjects | Pearson correlation (R): 0.5279 -> 0.9248 | Helped |
| Replace MEME with PNBA | Cross-subject alignment on PMd new subjects · PMd new subjects | Pearson correlation (R): 0.5255 -> 0.9176 | Helped |
| Replace MEME with PNBA | Cross-subject alignment on V1 training subjects · V1 training subjects | Pearson correlation (R): 0.6357 -> 0.8830 | Helped |
| Replace MEME with PNBA | Cross-subject alignment on V1 new subjects · V1 new subjects | Pearson correlation (R): 0.5980 -> 0.8705 | Helped |
| Add probabilistic matching to VAE modeling (ablation) | V1 alignment using only VAE · V1 neural-to-behavior alignment | Pearson correlation (R): 0.77 -> 0.87 | Helped |
| Add VAE modeling and probabilistic matching constraints (ablation) | V1 alignment using multi-modal VAE constraints · V1 neural-to-behavior alignment | Pearson correlation (R): 0.71 -> 0.87 | Helped |
| Set latent dimension to 32 (ablation) | PNBA on V1 · V1 neural-to-behavior alignment | Pearson correlation (R): R = 0.87 | Helped |
| Use higher-than-32 latent dimensionality (ablation) | PNBA on V1 · V1 neural-to-behavior alignment | Pearson correlation (R) | No clear effect |
| Use 4-dimensional latent representations (ablation) | PNBA on M1 and PMd · Motor cortical areas | Pearson correlation (R) | No clear effect |
| Use a GRU rather than a linear decoder | V1-guided movement decoding · V1-guided movement decoding | R²: 0.866 -> 0.888 | Helped |
Surprising findings
- Naive probabilistic matching and SigLIP-style alignment collapse neural embeddings to near-identical points, with no significant matched vs mismatched discrimination.
Limitations
- Discussion acknowledges difficulty extending to covert cognitive processes without observable behavior.
- Theoretical guarantees rely on heuristic arguments and assumed convergence of latent distributions.
Experimental setup
Datasets: Monkey M1/PMd center-out reaching recordings Mouse V1 two-photon calcium imaging with paired dynamic visual stimuli
Benchmarks: Monkey M1 center-out reaching Monkey PMd center-out reaching Mouse V1 calcium imaging
Metrics: single-trial Pearson correlation (R) explained variance (R^2) t-test significance
Baselines: VAE FA+Procrustes PCA+CCA FA+amLDS Neuroformer MEME
Builds on: MEME CLIP SigLIP variational autoencoder CEBRA
Closest papers by meaning
- Bridging Critical Gaps in Convergent Learning: How Representational Alignment Evolves Across Layers, Training, and Distribution Shifts (NeurIPS 2025 · similarity 0.88)
- Spectral Analysis of Representational Similarity with Limited Neurons (NeurIPS 2025 · similarity 0.87)
- Model–Behavior Alignment under Flexible Evaluation: When the Best-Fitting Model Isn’t the Right One (NeurIPS 2025 · similarity 0.87)
- Multi-Modal Latent Variables for Cross-Individual Primary Visual Cortex Modeling and Analysis (AAAI 2025 · similarity 0.86)
- SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding (NeurIPS 2025 · similarity 0.86)
- Sparse components distinguish visual pathways & their alignment to neural networks (ICLR 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.