Isolate the end-to-end effect of mean-centering V-JEPA-2 targets

Establish the causal effect of mean-centering V-JEPA-2 video embeddings on EEG–video alignment by conducting a controlled end-to-end ablation with learning rate and epoch budget held fixed between centered and uncentered training.

Background

The paper argues that mean-centering expands the discriminative geometry of V-JEPA-2 embeddings and prevents the teacher distribution from becoming nearly uniform. The geometric evidence supports this mechanism, but the reported end-to-end comparison changes both the learning rate and the number of epochs in addition to changing whether targets are centered.

Because those optimization changes are confounded with centering, the quantitative improvement attributed to centering has not been causally isolated. A fixed-learning-rate, fixed-budget ablation would determine whether centering itself produces the observed alignment gain.

References

We should be clear about what is not established. The evidence above is geometric, measured on the target space directly. There is no controlled end-to-end ablation of centering at fixed learning rate and epoch budget: the comparison sometimes quoted for it (a centered run at $+0.0272$ mean $\Delta R2$ versus an uncentered one at $+0.0014$) also changes the learning rate ($3{\times}10{-4}!\to!10{-4}$) and the epoch count ($50!\to!499$), so it cannot carry the attribution.

— Scaling subjects in cross-modal alignment: video decoding with EEG foundation model  (2610.09287 - Truong et al., 7 Oct 2026) in Appendix, Section “Why the V-JEPA-2 targets are mean-centered”