Generalization gap prediction on MNIST benchmark leaderboard · SOTA2 Research

Generalization gap prediction on MNIST

0.551Kendall CorrelationFourier generalization measure

Evaluation Results

Method	Links
Fourier generalization measure2026.06		0.551
PAC-Bayes flatness2026.06		0.546
PAC-Bayes init2026.06		0.516
PAC-Bayes origin2026.06		0.516
inverse margin2026.06		0.386
PAC-Bayes magnitude flatness2026.06		0.364
PAC-Bayes magnitude init2026.06		0.111
PAC-Bayes magnitude origin2026.06		0.046
path norm margin2026.06		-0.02
path norm2026.06		-0.046
distance spectral init2026.06		-0.085
Weight watcher2026.06		-0.218
log-prod Frobenius margin2026.06		-0.242
log-sum Frobenius margin2026.06		-0.242
log-prod Frobenius2026.06		-0.268
log-sum Frobenius2026.06		-0.268
L2-distance2026.06		-0.294
Frobenius distance2026.06		-0.294
L2-norm2026.06		-0.307
parameter norm2026.06		-0.307
# of parameters2026.06		-0.327
log-prod spectral2026.06		-0.333
log-prod spectral margin2026.06		-0.333
log spectral init2026.06		-0.333
Frobenius spectral2026.06		-0.333
log spectral origin2026.06		-0.333
log-sum spectral margin2026.06		-0.333
log-sum spectral2026.06		-0.333