Learning Dynamic Latent Spaces for Lifelong Generative Modelling

Task Free Continual Learning (TFCL) aims to capture novel concepts from non-stationary data streams without forgetting previously learned knowledge. Mixture models, which add new components when certain conditions are met, have shown promising results in TFCL tasks. However, such approaches do not make use of the knowledge already accumulated for positive transfer. In this paper, we develop a model, namely Online Recursive Variational Autoencoder (ORVAE). ORVAE utilizes prior by selectively incorporating newly learnt information, adding components, according known past data. We introduce attention mechanism regularize structural latent space most important information is reused while that interferes with samples inactivated. The proposed can maximize benefit forward transfer learning perform several experiments show achieves state-of-the-art under TFCL.