Meta-Learned Attribute Self-Gating for Continual Generalized Zero-Shot Learning

Zero-shot learning (ZSL) has been shown to be a promising approach to\ngeneralizing a model to categories unseen during training by leveraging class\nattributes, but challenges still remain. Recently, methods using generative\nmodels to combat bias towards classes seen during training have pushed the\nstate of the art of ZSL, but these generative models can be slow or\ncomputationally expensive to train. Additionally, while many previous ZSL\nmethods assume a one-time adaptation to unseen classes, in reality, the world\nis always changing, necessitating a constant adjustment for deployed models.\nModels unprepared to handle a sequential stream of data are likely to\nexperience catastrophic forgetting. We propose a meta-continual zero-shot\nlearning (MCZSL) approach to address both these issues. In particular, by\npairing self-gating of attributes and scaled class normalization with\nmeta-learning based training, we are able to outperform state-of-the-art\nresults while being able to train our models substantially faster\n($>100\\times$) than expensive generative-based approaches. We demonstrate this\nby performing experiments on five standard ZSL datasets (CUB, aPY, AWA1, AWA2\nand SUN) in both generalized zero-shot learning and generalized continual\nzero-shot learning settings.\n