Trainable Reference Vectors Accelerate Neural Network Early Exits With Minimal Overhead
Researchers demonstrate that replacing BranchyNet classifier heads with trainable class reference vectors accelerates ResNet-50 inference by up to 1.67 times ...
Researchers demonstrate that replacing BranchyNet classifier heads with trainable class reference vectors accelerates ResNet-50 inference by up to 1.67 times ...
A federated learning framework combining feature distillation and momentum aggregation reduced communication volume by up to 97.6 percent while improving ...
Researchers at Shanghai Polytechnic University have developed CDAFIM, a class-difficulty-aware incentive mechanism that uses Shapley values to fairly reward federated ...
© 2026 Scienmag - Science Magazine
© 2026 Scienmag - Science Magazine