Invariant Meta Learning for Out-of-Distribution Generalization
This addresses a limitation in meta-learning for researchers and practitioners needing robust few-shot learning, but it appears incremental as it builds on existing meta-learning with regularization for OOD scenarios.
The paper tackles the problem of meta-learning failing under out-of-distribution (OOD) tasks by proposing invariant meta learning, which finds invariant meta-initializations and adapts quickly with regularization, showing effectiveness in experiments on OOD few-shot tasks.
Modern deep learning techniques have illustrated their excellent capabilities in many areas, but relies on large training data. Optimization-based meta-learning train a model on a variety tasks, such that it can solve new learning tasks using only a small number of training samples.However, these methods assumes that training and test dataare identically and independently distributed. To overcome such limitation, in this paper, we propose invariant meta learning for out-of-distribution tasks. Specifically, invariant meta learning find invariant optimal meta-initialization,and fast adapt to out-of-distribution tasks with regularization penalty. Extensive experiments demonstrate the effectiveness of our proposed invariant meta learning on out-of-distribution few-shot tasks.