Bayesian compression for dynamically expandable networks

Abstract

This paper develops Bayesian Compression for Dynamically Expandable Network (BCDEN), which can learn a compact model structure with preserving the accuracy in a continual learning scenarios. Dynamically Expandable Network (DEN) is efficiently trained by performing selective retraining, dynamically expands network capacity with only the necessary number of units, and effectively prevents semantic drift by duplicating and timestamping units in an online manner. Overcoming conventional DEN only giving point estimates, we providing the Bayesian inference under the principle framework. We validate our BCDEN on multiple public datasets under continual learning setting, on which it can outperform existing continual learning methods on a variety of tasks, and with the state-of-the-art compression results, while still maintaining comparable performance.