The paper introduces ‘Adapt & Align’, a method for continual learning of neural networks using generative models. The method addresses the issue of performance loss in neural networks when retrained with additional data from different distributions. The proposed method splits the update process into two parts. Firstly, a local generative model is trained using only data from a new task. Secondly, the latent representations from the local model are consolidated with a global model that encodes knowledge of all past experiences. The paper demonstrates the application of this method with Variational Autoencoders (VAE) and Generative Adversarial Networks (GAN).
Publication date: 21 Dec 2023
Project Page: https://arxiv.org/abs/2312.13699
Paper: https://arxiv.org/pdf/2312.13699
