arXiv 1710.10628
Variational Continual Learning
By Cuong V. Nguyen, Yingzhen Li, et al.
Published 2017-10-29
Discussion
Read the public discussion and references gathered around this paper.
This paper develops variational continual learning (VCL), a simple but general framework for continual learning that fuses online variational inference (VI) and recent advances in Monte Carlo VI for neural networks. The framework can successfully train both deep discriminative models and deep generative models in complex continual learning settings where existing tasks evolve over time and entirely new tasks emerge.…