Research map / Probabilistic and scientific modeling / Dynamical systems
Neuron Platonic Intrinsic Representation From Dynamics Using Contrastive Learning
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 NeurPIR, which applies VICReg contrastive learning to multi-segment neuron activity (with CEBRA-encoded peripheral information) to learn time-invariant neuron embeddings that reflect intrinsic properties such as type and brain location.
Problem
Learning time-invariant, intrinsic neuron representations from neuronal activity recordings so that they predict neuron identity properties (type, molecular profile, brain region) and generalize to unseen animals.
Approach
Treat recording segments from the same neuron as positive pairs and segments from different neurons as negatives, encode surrounding information with CEBRA, and train a VICReg-based encoder so that only positive pairs are needed while variance and covariance regularizers keep neurons separated.
Key results
- On simulated Izhikevich data, NeurPIR reaches F1 0.884 for regular spiking and 0.881 for fast spiking, the best among compared methods.
- On Bugeon, NeurPIR gets class-level F1 0.747 vs 0.706 for NeuPRINT and 0.552 for LOLCAT.
- On Steinmetz brain-region classification, NeurPIR's out-of-domain (unseen animal) F1 is 0.708 vs 0.662 for NeuPRINT and 0.652 for LOLCAT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NeurPIR representations instead of the strongest listed baseline | Izhikevich simulation neuron-type classification · RS | F1: 0.870 -> 0.884 | Helped |
| Use NeurPIR representations instead of the strongest listed baseline | Izhikevich simulation neuron-type classification · IB | F1: 0.644 -> 0.684 | Helped |
| Use NeurPIR representations instead of the strongest listed baseline | Izhikevich simulation neuron-type classification · CH | F1: 0.652 -> 0.671 | Helped |
| Use NeurPIR representations instead of the strongest listed baseline | Izhikevich simulation neuron-type classification · FS | F1: 0.823 -> 0.881 | Helped |
| Use NeurPIR representations instead of the strongest listed baseline | Izhikevich simulation neuron-type classification · LTS | F1: 0.988 -> 0.990 | Helped |
| Use NeurPIR instead of NeuPRINT | Bugeon neuron subclass classification · Bugeon | F1: 0.553 ± 0.080 -> 0.582 ± 0.059 | Helped |
| Use NeurPIR instead of NeuPRINT | Bugeon neuron class classification · Bugeon | F1: 0.706 ± 0.047 -> 0.747 ± 0.016 | Helped |
| Use NeurPIR instead of NeuPRINT | Steinmetz brain-region classification · In-domain | F1: 0.743 ± 0.071 -> 0.758 ± 0.066 | Helped |
| Use NeurPIR instead of LOLCAT | Steinmetz brain-region classification · Out-of-domain | F1: 0.652 ± 0.073 -> 0.708 ± 0.054 | Helped |
| Replace NeurPIR with PCA | Izhikevich simulation neuron-type classification · Across firing modes | F1: RS: 0.786 -> 0.884; IB: 0.440 -> 0.684; CH: 0.548 -> 0.671; FS: 0.689 -> 0.881; LTS: 0.952 -> 0.990 | Mixed |
| Replace NeurPIR with UMAP | Izhikevich simulation neuron-type classification · Across firing modes | F1: RS: 0.525 -> 0.884; IB: 0.320 -> 0.684; CH: 0.285 -> 0.671; FS: 0.618 -> 0.881; LTS: 0.954 -> 0.990 | Mixed |
Surprising findings
- NeuPRINT achieves the highest precision (0.704) for the Sst subclass, though with a large standard deviation, and NeuPRINT is competitive on the LTS simulated class.
Limitations
- Can only distinguish neurons with substantial differences in essential attributes; fine-grained brain-area labels are hard to separate.
- Learned representations only generalize within the same recording technology; cross-platform (two-photon to Neuropixels) generalization is unexplored.
- Over very long timescales, some short-term invariant properties may change.
Experimental setup
Datasets: Izhikevich simulated neuron population data Bugeon two-photon and spatial transcriptomics dataset Steinmetz Neuropixels dataset
Benchmarks: Izhikevich simulated firing-mode classification Bugeon neuron type classification (subclass and class) Steinmetz brain region classification (in-domain and out-of-domain)
Metrics: precision recall F1
Baselines: PCA UMAP NeuPRINT LOLCAT
Builds on: VICReg CEBRA Izhikevich model UMAP NeuPRINT
Closest papers by meaning
- NOBLE - Neural Operator with Biologically-informed Latent Embeddings to Capture Experimental Variability in Biological Neuron Models (NeurIPS 2025 · similarity 0.87)
- Reproducibility of predictive networks for mouse visual cortex (NeurIPS 2024 · similarity 0.87)
- Neural Representational Consistency Emerges from Probabilistic Neural-Behavioral Representation Alignment (ICML 2025 · similarity 0.85)
- Time-Evolving Dynamical System for Learning Latent Representations of Mouse Visual Neural Activity (NeurIPS 2025 · similarity 0.85)
- Know Thyself by Knowing Others: Learning Neuron Identity from Population Context (NeurIPS 2025 · similarity 0.85)
- Spectral Analysis of Representational Similarity with Limited Neurons (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.