A finite-bank MNIST construction shows that an empirical estimator of algorithmic mutual information I(S;W) monotonically tracks the generalization gap in MoE models while providing an accuracy-rate curve via Blahut-Arimoto.
Mutual information neural estimation,
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Expert Routing for Communication-Efficient MoE via Finite Expert Banks
A finite-bank MNIST construction shows that an empirical estimator of algorithmic mutual information I(S;W) monotonically tracks the generalization gap in MoE models while providing an accuracy-rate curve via Blahut-Arimoto.