Skip to content

itsShnik/internal-knowledge-distillation

Repository files navigation

Internal Knowledge Distillation

Residual networks can be viewed as the ensembles of a lot of shallower sub-networks. The idea is to train the residual networks in such a way that the knowledge in the ensemble is distilled into the sub-networks in a single procedure. The advantages of doing the same are

  1. Increment in the accuracy of the original ResNet
  2. Possible training of residual networks of multiple depths in a single and efficient procedure
  3. A better approach for knowledge distillation when compared to the traditional distillation methods.

About

Internal knowledge distillation in residual networks

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages