A Nested Bi-level Optimization Framework for Robust Few Shot Learning

Model-Agnostic Meta-Learning (MAML), a popular gradient-based meta-learning framework, assumes that the contribution of each task or instance to meta-learner is equal.Hence, it fails address domain shift between base and novel classes in few-shot learning. In this work, we propose robust algorithm, NESTEDMAML, which learns assign weights training tasks instances. We con-sider as hyper-parameters iteratively optimize them using small set validation nested bi-level optimization approach (in contrast standard MAML). then applyNESTED-MAMLin meta-training stage, involves (1) several sampled from distribution different meta-test distribution, (2) some data samples with noisy labels.Extensive experiments on synthetic real-world datasets demonstrate NESTEDMAML efficiently mitigates effects ”unwanted” instances, leading significant improvement over state-of-the-art methods.