Continual Learning with Generative Replay via Discriminative Variational Autoencoder

Catastrophic forgetting in machine learning is the problem of gradually losing accuracy from previously learned tasks as new tasks are learned sequentially. This problem occurs mainly in modern artificial neural networks using gradient-based training algorithms. In this paper, we suggest a simple but robust generative replaybased model to mitigate the catastrophic forgetting problem. Also, We categorize current methods for incremental learning task as two approaches. Then, we analyze our method on Permuted MNIST task and Split MNIST task. Our experimental results show that our proposed method achieves competitive accuracy compared to other algorithms in Permuted MNIST task and outperforms other algorithms on Split MNIST task.