Deep online classification using pseudo-generative models

Andrey Besedin, Pierre Blanchart, Michel Crucianu, Marin Ferecatu

Abstract: Highlights•When learning in data streams, performance strongly depends on missing data.•Problems in non IID streams may be alleviated by using generative models.•Auto-encoders can be used to compactly store historical data.•Auto-encoders can replace fully fledged generative models for online learning.