Continual Learning from the Perspective of Compression
This work addresses the problem of catastrophic forgetting for researchers and practitioners in continual learning, offering a novel theoretical perspective and a hybrid method, though it is incremental in building on existing compression-based approaches.
The paper tackles catastrophic forgetting in neural networks by analyzing it through information theory, defining forgetting as increased description lengths of previous data when compressed with a sequentially learned model, and proposes a new method combining ML plug-in and Bayesian mixture codes, showing empirical improvements in compression and forgetting metrics.
Connectionist models such as neural networks suffer from catastrophic forgetting. In this work, we study this problem from the perspective of information theory and define forgetting as the increase of description lengths of previous data when they are compressed with a sequentially learned model. In addition, we show that continual learning approaches based on variational posterior approximation and generative replay can be considered as approximations to two prequential coding methods in compression, namely, the Bayesian mixture code and maximum likelihood (ML) plug-in code. We compare these approaches in terms of both compression and forgetting and empirically study the reasons that limit the performance of continual learning methods based on variational posterior approximation. To address these limitations, we propose a new continual learning method that combines ML plug-in and Bayesian mixture codes.