Historically, applying classical PAC-Bayes theory to overparameterized models resulted in loose a..., Sonic AI
“Historically, applying classical PAC-Bayes theory to overparameterized models resulted in loose and unusable bounds because the compression term would dominate.”