LFPformer

Geometry-Aware Self-Supervised Electrophysiology Representation Learning
NeurIPS 2026
S. DeyP.-J. LuV. S. ParupudiM. PatelM.-E. Conde Paredes M. VöröslakosA. MaslarovaG. BuzsákiS. VyasE. Varol

Self-supervised learning on raw LFP that places every Neuropixels channel in the Allen brain atlas.

3D explorer
Loading the Allen brain and probe tracks…
Ground truth Prediction
Method
Predicted track
Probes

IBL test probes, full recordings. Drag to rotate, click a track.

Highlights

Localization
~1 mm per probe
LFP only: 1012 µm vs 1484 (AnyNet), 1576 (ViT)
Fusion
LFP + micromanipulator
710 µm vs 790 µm for the micromanipulator alone
Pretraining
Anatomy emerges
Atlas coordinates appear in frozen embeddings
Scale
257M LFP chunks
95M and 318M encoders
Transfer
Region decoding
309 IBL regions: 0.081 vs 0.035 balanced acc.

Prediction vs ground truth

One tile per test probe. Coloured line: truth · dots: pooled prediction · dark line: Kabsch-aligned.

Sort
Section

Chronic recording, zero-shot

One mouse, Neuropixels 2.0, five days over 2.5 months. No histology.

Full session

Loading…
Regions along the probe
surface
tip

Linear maze

Mouse position
Regions along the probe
surface
tip

From LFP to atlas coordinates

LFPformer readouts

Results

From the paper (Fig. 3c, Fig. 4).

IBL 3D localization
Per-probe 3D error ↓
LFP only
With micromanipulator coordinates
−472 µm vs AnyNet · −565 µm vs ViT
Brain-region classification
Balanced accuracy ↑
IBL · 309 leaf regions (chance 0.003)
NeuroNexus · 5 regions (chance 0.20)
2.3× AnyNet on 309 IBL regions
Scale
Data and models
Pretraining data
257M
3-s LFP chunks
Encoders
95M / 318M
base / large encoders
Test set
8.88M
test chunks · 119 probes

Pretrained models

Four public Large encoders, each trained for 35,000 steps. Extract 1,024-dimensional embeddings from 3-second LFP windows.

Get started: load a model and extract embeddings

CCFDistc uses atlas supervision during pretraining. Localization requires a separately trained readout.

Figures

BibTeX

@inproceedings{dey2026geometry,
  title     = {Geometry-Aware Self-Supervised Electrophysiology Representation Learning},
  author    = {Dey, S. and Lu, P.-J. and Parupudi, V. S. and Patel, M. and Conde Paredes, M.-E. and
               V{\"o}r{\"o}slakos, M. and Maslarova, A. and Buzs{\'a}ki, G. and Vyas, S. and Varol, E.},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}
Data: IBL Brain-Wide Map, Allen Brain Observatory; brain mesh from the Allen CCFv3.
Page layout adapted from the MARCO project page.