Incremental Meta-Learning via Episodic Replay Distillation for Few-Shot Image Recognition

Most meta-learning approaches assume the existence of a very large set of labeled data available for episodic metalearning of base knowledge. This contrasts with the more realistic continual learning paradigm in which data arrives incrementally in the form of tasks containing disjoint classes. In this paper we consider this problem of Incremental MetaLearning (IML) in which classes are presented incrementally in discrete tasks. We propose an approach to IML, which we call Episodic Replay Distillation (ERD), that mixes classes from the current task with class exemplars from previous tasks when sampling episodes for meta-learning. These episodes are then used for knowledge distillation to minimize catastrophic forgetting. Experiments on four datasets demonstrate that ERD surpasses the state-of-the-art. In particular, on the more challenging one-shot, long task sequence incremental meta-learning scenarios, we reduce the gap between IML and the joint-training upper bound from 3.5% / 10.1% / 13.4% with the current state-of-the-art to 2.6% / 2.9% / 5.0% with our method on Tiered-ImageNet / Mini-ImageNet / CIFAR100, respectively. Introduction Meta-learning, also commonly referred to as “learning to learn”, is a learning paradigm in which a model gains experience over a sequence of learning episodes.1 This experience is optimized so as to improve the model’s future learning performance on unseen tasks (Hospedales et al. 2021). Meta-learning is one of the most promising techniques to learning models that can flexibly generalize, like humans, to new tasks and environments not seen during training. This capability is generally considered to be crucial for future AI systems. Few-shot learning has emerged as the paradigm-ofchoice to test and evaluate meta-learning algorithms. It aims to learn from very limited numbers of samples (as few as just one), and meta-learning applied to few-shot image recognition in particular has attracted increased attention in recent years (Su, Maji, and Hariharan 2020; Bateni et al. 2020; Li et al. 2020; Yang, Liu, and Xu 2021). However, most few-shot learning methods are limited in their learning modes: they must train with a large number of To avoid ambiguities, we use the term episode in the sense used in meta-learning rather than how it is used in continual learning. We use task in the sense of continual learning to refer to a disjoint group of new classes. Meta-test data Episode 1 Episode 2 Episode n ··· Model (init) Exem. mem. (empty) Episode 3 Model (t=3) Task 3 data