CMVAE: Causal Meta VAE for Unsupervised Meta-Learning

Unsupervised meta-learning aims to learn the meta knowledge from unlabeled data and rapidly adapt novel tasks. However, existing approaches may be misled by context-bias (e.g. background) training data. In this paper, we abstract unsupervised problem into a Structural Causal Model (SCM) point out that such bias arises due hidden confounders. To eliminate confounders, define priors are conditionally independent, relationships between intervene on them with casual factorization. Furthermore, propose Meta VAE (CMVAE) encodes latent codes in causal space learns their simultaneously achieve downstream few-shot image classification task. Results toy datasets three benchmark demonstrate our method can remove it outperforms other state-of-the-art algorithms because of bias-removal. Code is available at https://github.com/GuodongQi/CMVAE.