Temporal Alignment in MEG

Abstract

Does alignment complexity mirror the cortical hierarchy?

We test whether the complexity of cross-subject neural alignment increases with post-stimulus time in MEG — the hypothesis that early visual representations align via simple rigid rotations while late semantic representations require more flexible transforms.

Using THINGS-MEG (4 subjects, 272-channel CTF MEG, 1,854 object images) and a large-scale EEG replication (THINGS-EEG1, 48 subjects), we find that alignment complexity is flat across all timepoints and all PCA dimensionalities — definitively rejecting the hypothesis. A temporal generalization analysis reveals the mechanism: the geometric relationship between subjects is time-invariant, meaning a map trained at 20 ms transfers equally to 400 ms.

A systematic brain-DNN comparison reveals a clear supervision gradient: architecture alone (ViT-B/16 supervised ≈ random ViT-B/32) contributes nothing; joint visual-language training (CLIP) is required for full alignment. Source-space RSA confirms the alignment peaks in ventral stream regions (temporal pole, parahippocampal cortex), consistent with the CLIP model's semantic rather than purely visual representations.

TL;DR: Early-simple/late-complex alignment is rejected across MEG and EEG. Cross-subject geometry is time-invariant. CLIP's brain alignment requires joint visual-language training — architecture and ImageNet classification alone are insufficient.

Key Results

Findings at a glance

n=52

Total subjects across MEG (n=4) and EEG replication (n=48)

1,854

Object concepts — full THINGS stimulus set

81.5%

Cross-subject alignment as % of within-subject noise ceiling (EEG)

p<.001

Cluster permutation test confirming complexity null (n=48)

4.2×

CLIP-image vs. random ViT-B/32 brain RSA ratio

r=0.218

Peak source-space RSA — temporal pole at 495 ms (CLIP)

Primary finding — rejected hypothesis Alignment complexity (ridge − Procrustes accuracy) is flat and slightly negative across all 180 timepoints in both MEG (n=4) and EEG (n=48, p < 0.001). A temporal generalization matrix shows maps trained at 20 ms transfer equally to 400 ms — the cross-subject geometric transform is time-invariant.

Secondary finding — supervision gradient A clear hierarchy: ViT-B/16 supervised ≈ random ViT-B/32 (r ≈ 0.008–0.009) < CLIP-text (r = 0.028) < CLIP-image (r = 0.043). Architecture alone contributes nothing. Language-grounded visual training is required for full brain alignment.

Spatial finding — ventral stream Source-space RSA identifies temporal pole and entorhinal cortex as peak regions (r = 0.218, 0.214 at ~495 ms), with lateral occipital cortex peaking earlier (580 ms) — consistent with the ventral visual stream hierarchy.

Brain-DNN RSA

Model comparison across 1,852 categories

All models tested against the full crossnobis RDM at each of 180 timepoints. Expressed as peak Spearman r and percentage of noise ceiling.

Model	Training	Peak r	% Ceiling	Peak time
SPOSE	Human similarity judgments	0.046	17.3%	380 ms
CLIP ViT-B/32	Image + language (400M pairs)	0.043	16.1%	325 ms
CLIP ViT-L/14	Image + language (400M pairs)	0.030	11.2%	340 ms
ResNet-50	ImageNet classification	0.029	10.9%	410 ms
CLIP-text	Text encoder only	0.028	10.5%	410 ms
DINOv2 ViT-B/14	Self-supervised	0.013	4.8%	380 ms
ViT-B/16 supervised	ImageNet classification	0.008	3.0%	390 ms
Random ViT-B/32	Untrained (control)	0.009	3.4%	—

DINOv2 anomaly DINOv2 anti-correlates with brain responses for animate objects specifically, winning only on texture-heavy categories. Its self-supervised objective produces texture-based groupings that diverge from the semantic structure of human neural representations.

Methods

Analysis pipeline

1

MEG preprocessing

MNE-Python · bandpass 0.1–40 Hz · infomax ICA (manual verification) · epoch −100–800 ms · baseline · 200 Hz · condition-average 12 trials/image

2

Crossnobis RDMs

Cross-validated Mahalanobis distance (Walther et al. 2016) with Ledoit-Wolf precision matrix · computed per subject per timepoint · full 1,854-category RDMs

3

Cross-subject alignment

Procrustes (orthogonal, rigid rotation) vs. ridge regression (unconstrained linear) · 5-fold CV · complexity = ridge − Procrustes accuracy at each timepoint · EEG replication n=48

4

Brain-DNN RSA

Spearman r between brain crossnobis RDM and model RDMs at each timepoint · 5 models + 3 controls · FDR correction (Benjamini-Hochberg) · noise ceiling (Nili et al. 2014) · partial RSA for unique contributions

5

Source localization

MNE minimum-norm inverse · fsaverage template · ico-4 (5,124 sources) · Desikan-Killiany parcellation (68 regions) · RSA per parcel × timepoint

6

Controls

Cluster permutation testing (1,000 permutations) · alignment noise ceiling · PCA dimensionality robustness (k=5–63) · temporal generalization matrix · random-weights baseline · category decomposition

Citation

BibTeX

@article{seghir2026meg,
  title   = {Temporal Dynamics of Cross-Subject Representational Alignment in Human MEG},
  author  = {Seghir, Anouar},
  year    = {2026},
  doi     = {10.17605/OSF.IO/JZFSA},
  url     = {https://doi.org/10.17605/OSF.IO/JZFSA},
  note    = {OSF Preprint}
}