ABKD replaces forward and reverse KL distillation with a two-parameter alpha-beta divergence that interpolates between them, improving results on 17 language and vision benchmarks.
Ify1 andy2 are such thatδ1 <q t(y1) =q t(y2)≤p(y 1) (whereδ1 >0 ), andp(y1)≥p(y 2) +ζ , it holds that∆α1 t (y1,y 2)≥∆ α2 t (y1,y 2)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
ABKD: Pursuing a Proper Allocation of the Probability Mass in Knowledge Distillation via $\alpha$-$\beta$-Divergence
ABKD replaces forward and reverse KL distillation with a two-parameter alpha-beta divergence that interpolates between them, improving results on 17 language and vision benchmarks.