#18 · ML / AI methods

Brain-model alignment is probed via gradients, training dynamics, and EEG decoding

The dominant thread is diagnostic evaluation of how computational representations correspond to human neural data, rather than headline predictive accuracy. Gradient-based mapping of DINOv3 and other vision models onto fMRI and MEG finds that backpropagated gradients mismatch the hierarchy of brain responses, asking where model and brain hierarchies diverge. A training study reports that supervised training degrades RSA alignment to human V1 fMRI across all tested biologically plausible learning rules, with an earlier PC/STDP preservation claim corrected as unsupported. EEG-FM-Audit proposes an evaluation pipeline with tuned supervised baselines and neurophysiological probing for EEG foundation models, and a co-skewness analysis finds brain foundation models lose third-order fMRI statistics without outperforming FC regression on cognition. MindAlign, by contrast, is primarily performance-oriented, reaching 54.1% zero-shot Top-1 on Things-EEG2 200-way decoding.