Towards lifelong learning for generative adversarial networks

Humans are lifelong learners, we acquire and accumulate knowledge throughout our lives. In contrast to human learning, neural networks are susceptible to catastrophic forgetting: when adapted to perform new tasks, they often fail to preserve their performance on previously learned tasks. Lifelong learning ability is crucial for the development of artificial intelligence, and many practical vision applications require such continual learning ability. Continual learning has been well explored for discriminative tasks, i.e. classification. Compared with discriminative models, lifelong learning for generative models remains an important yet under-explored area. In this dissertation, we explore the applications of generative models, study the problem of lifelong learning for generative models and propose novel generative lifelong learning algorithms. In detail, we present (1) the first generic lifelong learning method enabling both label-conditioned and image-conditioned generation tasks; (2) a parameter efficient lifelong learning algorithm which does not suffer from performance degradation; (3) a scalable lifelong learning method that is parameter efficient and quality-preserving.