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.
q(y) 1−α (p(y)α−q(y) α) +q(y) X k q(k) 1−α(q(k)α−p(k)α) !#! .(47) Taking the logarithm on both sides, we get: logqα t+1(y) qt(y) =η
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.