Research map / Time series / Foundation models for time series
TRACE: Contrastive learning for multi-trial time series data in neuroscience
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 TRACE, a contrastive learning framework that builds positive pairs from averages of random non-overlapping trial subsets and directly learns 2D embeddings of multi-trial neural time series using a Cauchy similarity.
Problem
Visualizing and clustering large multi-trial neural time-series datasets (e.g., two-photon imaging, Neuropixels) in 2D; existing contrastive methods use generic augmentations and do not embed directly into 2D.
Approach
Create positive pairs as means of two non-overlapping random subsets of trials of the same neuron (mix-up-like, trial-based augmentation), and train a parametric encoder directly to 2D with a t-SNE-style Cauchy kernel InfoNCE loss, building on t-SimCNE.
Key results
- On the superior colliculus dataset, TRACE achieves the best ARI (0.28) and average rank (2.17), with kNN accuracy 69.7% vs. 70.7% for t-SNE.
- TRACE ranks best on the Neuropixels dataset (average rank 1.8) with kNN accuracy 23.7% and behavior correlation 0.11, outperforming CEED in all metrics.
- On synthetic data at high baseline noise, TRACE keeps discriminability (3.52 at noise SD 60) while CEED, t-SNE, and UMAP drop to below 0.2 to 0.05 range for several noise levels.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TRACE rather than competing visualization methods | Synthetic responses · Synthetic responses, noise SD 60 | mean discriminability: 3.52 vs 0.18 (CEED → PCA), 0.05 (TS2Vec → PCA), 0.03 (t-SNE), 0.05 (UMAP) | Helped |
| Add trial-subset mean positive pairs instead of CEED augmentations (ablation) | TRACE · Superior colliculus calcium imaging | ARI: 0.25 -> 0.28 | Helped |
| Add trial-subset mean positive pairs instead of TS2Vec approach (ablation) | TRACE · Superior colliculus calcium imaging | ARI: 0.10 -> 0.28 | Helped |
| Replace trial-subset means with TS2Vec approach (ablation) | TRACE · Superior colliculus calcium imaging | Spearman correlation: 0.45 ± 0.01 -> 0.60 ± 0.01 | Hurt |
| Use TRACE rather than t-SNE | 2D embedding · Superior colliculus calcium imaging | ARI: 0.24 -> 0.28 | Helped |
| Use TRACE rather than t-SNE | 2D embedding · Superior colliculus calcium imaging | kNN accuracy: 69.7 ± 0.3% vs 70.7 ± 0.2% | Mixed |
| Replace trial-subset means with CEED augmentations (ablation) | TRACE · Neuropixels | average rank: 1.8 -> 3.5 | Hurt |
| Replace trial-subset means with TS2Vec approach (ablation) | TRACE · Neuropixels | average rank: 1.8 -> 4.5 | Hurt |
| Use TRACE rather than CEED | 2D embedding · Neuropixels | ARI: 0.03 -> 0.04 | Helped |
| Use larger subset size k instead of k=1 (ablation) | TRACE · Superior colliculus calcium imaging, varying subset size k | kNN accuracy: 51.9 -> 69.5 | Helped |
Surprising findings
- TRACE still performs well with k=1 trial per subset in the calcium imaging dataset.
- t-SNE and UMAP fail on the synthetic data because distances are dominated by uninformative baseline noise.
Limitations
- Requires multi-trial (repeated stimulus) data; not applicable to single-trial recordings.
- Training an embedding model is slower than t-SNE or UMAP.
- Assumes between-trial variation is noise, so it is unsuited to studying trial-to-trial adaptation.
Experimental setup
Datasets: Two-photon calcium imaging of mouse retinal ganglion cell boutons in superior colliculus (71,021 boutons) Allen Institute Neuropixels visual coding dataset (10,322 neurons) Synthetic two-cell-type dataset
Benchmarks: Synthetic two-class dataset Superior colliculus calcium imaging dataset Allen Institute Neuropixels visual coding dataset
Metrics: ARI score kNN accuracy Spearman rank correlation linear/radial correlation with biological variables kNN regression R^2 average rank discriminability
Baselines: CEED TS2Vec t-SNE UMAP PCA
Builds on: t-SimCNE SimCLR CEED TS2Vec mix-up
Closest papers by meaning
- Inference via Interpolation: Contrastive Representations Provably Enable Planning and Inference (NeurIPS 2024 · similarity 0.86)
- Frequency-Masked Embedding Inference: A Non-Contrastive Approach for Time Series Representation Learning (AAAI 2025 · similarity 0.85)
- REBAR: Retrieval-Based Reconstruction for Time-series Contrastive Learning (ICLR 2024 · similarity 0.85)
- Contrastive Consolidation of Top-Down Modulations Achieves Sparsely Supervised Continual Learning (NeurIPS 2025 · similarity 0.85)
- Time-Evolving Dynamical System for Learning Latent Representations of Mouse Visual Neural Activity (NeurIPS 2025 · similarity 0.84)
- Contrastive Difference Predictive Coding (ICLR 2024 · 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.