Training multi-exit adaptive networks with gradient rescaling, inline logit sharing, and self-distillation improves their accuracy at fixed compute budgets.
Model compression
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2019 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Improved Techniques for Training Adaptive Deep Networks
Training multi-exit adaptive networks with gradient rescaling, inline logit sharing, and self-distillation improves their accuracy at fixed compute budgets.