Overcoming Catastrophic Forgetting by Bayesian Generative Regularization

The streaming update of Bayesian posterior calculation provides us a natural way for continual learning. However, the naïve mean-field posterior parametrization for variational approximation is inappropiate in neural network, and thus, lock the full ability for preventing catastrophic forgetting. To resolve this issue, we introduce a generative regularization for all given classification models, which is implemented by leveraging energy-based models with contrastive loss, to obtain the sufficient features for valid decomposition in posterior approxiamtion. By combining discriminative and generative loss together, we empirically show that the proposed method outperforms state-ofthe-art methods on a variety of tasks, avoiding catastrophic forgetting in continual learning. In particular, the proposed method outperforms baseline methods over 15% on the Fashion-MNIST dataset and 10% on the CUB dataset.