Hyper-LifelongGAN: Scalable Lifelong Learning for Image Conditioned Generation

Task 1 Task n-1 Task … Task n … weight matrix shared across all tasks random noise parameters needed for the next task Task 1 Task n-1 Task … Task n (a) Sequential Fine-tuning (b) Hyper-LifelongGAN base filter generator base filter generator base filter generator Figure 1: Lifelong learning of image-conditioned generation. Encountering a new task, traditional training methods forget how to perform previous tasks (Figure 1 (a)). Hyper-LifelongGAN is a scalable framework addressing catastrophic forgetting. It can adapt to the new task with few additional parameters, while preserving the knowledge of previous tasks(Figure 1 (b)).