Research map / Time series / Foundation models for time series
Population Transformer: Learning Population-level Representations of Neural Activity
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 Population Transformer (PopT), a self-supervised transformer that stacks on frozen per-channel temporal embeddings to learn population-level codes over variable electrode ensembles for iEEG and EEG decoding.
Problem
Decoding from multi-channel neural recordings is hard because electrode layouts are sparse and variable across subjects and datasets, so models must be retrained per participant with lots of calibration data.
Approach
Decouple temporal and spatial modeling: a transformer aggregates frozen single-channel temporal embeddings with 3D electrode position encodings, pretrained with discriminative ensemble-wise (consecutive-time) and channel-wise (swapped-channel) objectives instead of reconstruction.
Key results
- Pretrained PopT with BrainBERT reaches 0.74 pitch, 0.87 volume, 0.90 sentence onset and 0.93 speech/non-speech ROC-AUC on 90 iEEG electrodes, beating linear and deep NN aggregation and end-to-end Brant (0.61/0.74/0.80/0.80).
- On TUAB with Chronos embeddings, pretrained PopT reaches 0.7963 balanced accuracy and 0.8822 ROC AUC, versus 0.7881 and 0.8678 for Deep NN aggregation.
- PopT has about 20M parameters at 50 channels versus 350M for LaBraM and 500M for Brant; pretraining takes 2 days on one TITAN RTX.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| pretrain PopT on BrainBERT temporal embeddings rather than use end-to-end Brant | iEEG decoding · iEEG Pitch | ROC-AUC: 0.61 ± 0.03 -> 0.74 ± 0.03 | Helped |
| pretrain PopT on BrainBERT temporal embeddings rather than use end-to-end Brant | iEEG decoding · iEEG Volume | ROC-AUC: 0.74 ± 0.03 -> 0.87 ± 0.03 | Helped |
| pretrain PopT on BrainBERT temporal embeddings rather than use end-to-end Brant | iEEG decoding · iEEG Sentence Onset | ROC-AUC: 0.80 ± 0.04 -> 0.90 ± 0.01 | Helped |
| pretrain PopT on BrainBERT temporal embeddings rather than use end-to-end Brant | iEEG decoding · iEEG Speech/Non-speech | ROC-AUC: 0.80 ± 0.03 -> 0.93 ± 0.02 | Helped |
| pretrain PopT with Chronos embeddings rather than use LaBraM | TUAB abnormal EEG detection · TUAB | Balanced accuracy: 0.7963 ± 0.0021 vs 0.8258 ± 0.0011 | Hurt |
| pretrain PopT with TS2Vec embeddings rather than use LaBraM | TUAB abnormal EEG detection · TUAB | ROC AUC: 0.8883 ± 0.0008 vs 0.9162 ± 0.0016 | Hurt |
| add the channel-wise loss (ablation) | PopT · iEEG Pitch | ROC-AUC: 0.71 ± 0.03 -> 0.74 ± 0.03 | Helped |
| add the ensemble-wise loss (ablation) | PopT · iEEG Pitch | ROC-AUC: 0.70 ± 0.03 -> 0.74 ± 0.03 | Helped |
| add position encoding (ablation) | PopT · iEEG Pitch | ROC-AUC: 0.62 ± 0.07 -> 0.74 ± 0.03 | Helped |
| use discriminative losses rather than reconstruction loss only (ablation) | PopT · iEEG Pitch | ROC-AUC: 0.56 ± 0.04 -> 0.74 ± 0.03 | Helped |
| add Gaussian noise to coordinates before position encoding (ablation) | PopT · iEEG Pitch | ROC-AUC: 0.74 ± 0.03 -> 0.73 ± 0.03 | Hurt |
| replace a channel with its own activity at another time rather than another channel's activity (ablation) | PopT channel-wise pretraining · iEEG Pitch | ROC-AUC: 0.74 ± 0.03 -> 0.71 ± 0.02 | Hurt |
Surprising findings
- Non-pretrained PopT performs no better than or worse than simple linear and deep NN aggregation on several iEEG tasks, so pretraining is what drives the gains.
- Held-out-subject pretraining causes only a minimal drop in downstream decoding, suggesting generalization to unseen subjects.
Limitations
- Requires meaningful spatial coordinates for each electrode; unclear how to apply without them.
- Pretraining data is limited (19 movies, 10 subjects), and results show a plateau in scaling with current data.
- Connectivity and functional region analyses are qualitative and rely on attention and perturbation heuristics.
Experimental setup
Datasets: iEEG movie-watching dataset (10 subjects) TUEG TUAB
Benchmarks: iEEG audio-linguistic decoding (pitch, volume, sentence onset, speech/non-speech) TUAB abnormal EEG detection
Metrics: ROC-AUC balanced accuracy
Baselines: Linear aggregation Deep NN aggregation Non-pretrained PopT Brant BIOT LaBraM
Builds on: BrainBERT Brant BIOT LaBraM
Closest papers by meaning
- Neuroformer: Multimodal and Multitask Generative Pretraining for Brain Data (ICLR 2024 · similarity 0.88)
- SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding (NeurIPS 2025 · similarity 0.88)
- Identifying interactions across brain areas while accounting for individual-neuron dynamics with a Transformer-based variational autoencoder (NeurIPS 2025 · similarity 0.87)
- POCO: Scalable Neural Forecasting through Population Conditioning (NeurIPS 2025 · similarity 0.86)
- Towards a "Universal Translator" for Neural Dynamics at Single-Cell, Single-Spike Resolution (NeurIPS 2024 · similarity 0.86)
- Leveraging unlabelled data for generalizable neural population decoding (NeurIPS 2026 · 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.