Attention Consistency Regularization for Interpretable Early-Exit Neural Networks

Yanhua Zhao

Early-exit neural networks enable adaptive inference by allowing predictions at intermediate layers, reducing computational cost. However, early exits often lack interpretability a…