Lifelong Generative Modeling

Lifelong learning is the problem of learning multiple consecutive tasks in an online manner and is essential towards the development of intelligent machines that can adapt to their surroundings. In this work we focus on learning a lifelong approach to generative modeling whereby we continuously incorporate newly observed distributions into our model representation. We utilize two models, aptly named the student and the teacher, in order to aggregate information about all past distributions without the preservation of any of the past data or previous models. The teacher is utilized as a form of compressed memory in order to allow for the student model to learn over the past as well as present data. We demonstrate why a naive approach to lifelong generative modeling fails and introduce a regularizer with which we demonstrate learning across a long range of distributions.