Research map / Label-efficient and robust learning / Self-supervised and contrastive
Know Thyself by Knowing Others: Learning Neuron Identity from Population Context
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.
Introduces NuCLR, a self-supervised permutation-equivariant spatiotemporal transformer trained with a neuron-level contrastive objective to learn population-aware neuron identity embeddings that support zero-shot cell type and brain region decoding.
Problem
Inferring neuron identity (cell type, brain region) from large-scale in vivo population activity when anatomical or molecular labels are unavailable, with generalization across animals without retraining.
Approach
Contrastive learning at the neuron level: two temporal views of the same population form positives for the same neuron and all other neurons in the same insertion are negatives, with a permutation-equivariant spatiotemporal transformer providing population context and a decoupled contrastive loss computed per view pair.
Key results
- Inductive zero-shot Allen VC cell-type macro F1 0.7200 vs 0.4194 for NEMO (next best).
- Inductive zero-shot IBL brain-region macro F1 0.5295 vs 0.3793 for NEMO.
- Ablating spatial attention drops Allen VC macro F1 from 0.72 to 0.555.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove spatial attention layers (ablation) | NuCLR · Allen VC cell type, inductive zero-shot | macro F1: 0.7200 -> 0.5550 | Hurt |
| remove spatial attention layers (ablation) | NuCLR · IBL brain region, inductive zero-shot | macro F1: 0.5295 -> 0.3573 | Hurt |
| add neuron dropout (ablation) | NuCLR · Allen VC cell type, inductive zero-shot | macro F1: 0.6181 -> 0.7200 | Helped |
| add neuron dropout (ablation) | NuCLR · IBL brain region, inductive zero-shot | macro F1: 0.5248 -> 0.5295 | No clear effect |
| replace 20ms binning with spike-tokenization (ablation) | NuCLR · Allen VC cell type, inductive zero-shot | macro F1: 0.7200 -> 0.6665 | Hurt |
| replace 20ms binning with spike-tokenization (ablation) | NuCLR · IBL brain region, inductive zero-shot | macro F1: 0.5295 -> 0.5268 | No clear effect |
| replace temporal-only layers with spatio-temporal layers (ablation) | NuCLR · Allen VC cell type, inductive zero-shot | macro F1: 0.7200 -> 0.7184 | No clear effect |
| replace temporal-only layers with spatio-temporal layers (ablation) | NuCLR · IBL brain region, inductive zero-shot | macro F1: 0.5295 -> 0.5259 | No clear effect |
| increase the number of pretraining animals | NuCLR · Allen VC cell type and IBL brain region, inductive zero-shot | decoding performance | Helped |
| use only 12.5% of labeled data | NuCLR · Allen VC cell type, inductive zero-shot | score: 0.54 | Helped |
| use only 12.5% of labeled data | NuCLR · IBL brain region, inductive zero-shot | score: 0.50 | Helped |
| use NuCLR rather than the strongest reported baseline | Allen VC cell type, inductive zero-shot · Allen VC | macro F1: 0.4194 -> 0.7200 | Helped |
Surprising findings
- Spatial/population context is critical: removing spatial attention layers causes large drops on both Allen VC and IBL.
- Learned embeddings on small datasets (Bugeon, Steinmetz) cluster by subject or session identity rather than only by cell type or region.
Limitations
- Separate models are needed for electrophysiology and calcium imaging modalities.
- Small datasets yield embeddings that cluster by subject or session identity, entangling recording-specific factors.
- Evaluated only on datasets with similar task structure; robustness to stimulus distribution changes untested.
Experimental setup
Datasets: Allen Brain Observatory Visual Coding Neuropixels IBL Brain-wide Map Steinmetz et al. 2019 Bugeon et al. calcium imaging
Benchmarks: Allen VC IBL Brain-wide Map Steinmetz et al. 2019 Bugeon et al.
Metrics: macro F1-score
Baselines: NeuPRINT NEMO LOLCAT POYO+ NeurPIR
Builds on: SimCLR decoupled contrastive learning rotary position embeddings POYO+
Closest papers by meaning
- Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data (ICLR 2024 · similarity 0.86)
- BLEND: Behavior-guided Neural Population Dynamics Modeling via Privileged Knowledge Distillation (ICLR 2025 · similarity 0.86)
- Identifying interactions across brain areas while accounting for individual-neuron dynamics with a Transformer-based variational autoencoder (NeurIPS 2025 · similarity 0.86)
- Towards a "Universal Translator" for Neural Dynamics at Single-Cell, Single-Spike Resolution (NeurIPS 2024 · similarity 0.85)
- Neuron Platonic Intrinsic Representation From Dynamics Using Contrastive Learning (ICLR 2025 · similarity 0.85)
- NeurFlow: Interpreting Neural Networks through Neuron Groups and Functional Interactions (ICLR 2025 · similarity 0.84)
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.