Lifelong Learning in StyleGAN through Latent Subspaces
Abstract: StyleGAN is one of the most versatile generative models that have emerged in recent times. However, when it is trained continually on a stream of data (potentially previously unseen distributions), it tends to forget the distribution it has learned, as is the case with any other generative model, due to catastrophic forgetting. Recent studies have shown that the latent space of StyleGAN is very versatile, as data from a variety of distributions can be inverted onto it. In this paper, we propose to leverage this property to facilitate lifelong learning of StyleGAN without forgetting. Specifically, given a StyleGAN trained on a certain task (dataset), we propose to learn a latent subspace characterized by a set of dictionary vectors in its latent space, one for each novel, unseen task (or dataset). We also learn a relatively small set of parameters (feature adaptors) in the weight space to complement the dictionary learning in the latent space. Furthermore, we introduce a method that utilizes the similarity between tasks to effectively reuse the feature adaptor parameters from the previous tasks, aiding in the learning process for the current task at hand. Our approach guarantees that the parameters from previous tasks are reused only if they contribute to a beneficial forward transfer of knowledge. Remarkably, StyleCL avoids catastrophic forgetting because the set of dictionary and the feature adaptor parameters are unique for each task. We demonstrate that our method, StyleCL, achieves better generation quality on multiple datasets with significantly fewer additional parameters per task compared to previous methods. This is a consequence of learning task-specific dictionaries in the latent space, which has a much lower dimensionality compared to the weight space.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~ERIC_EATON1
Submission Number: 2561
Loading