Interventional Few-Shot Learning

TL;DR

This paper identifies a limitation in current Few-Shot Learning methods caused by pre-trained knowledge and introduces Interventional Few-Shot Learning (IFSL), a causal intervention approach that improves state-of-the-art results across multiple benchmarks.

Contribution

The paper proposes a novel causal intervention framework for FSL, called IFSL, which enhances existing methods and achieves new state-of-the-art performance.

Findings

IFSL improves 1-/5-shot performance on miniImageNet and tieredImageNet.

IFSL achieves state-of-the-art results on cross-domain CUB.

The approach is compatible with existing fine-tuning and meta-learning methods.

Abstract

We uncover an ever-overlooked deficiency in the prevailing Few-Shot Learning (FSL) methods: the pre-trained knowledge is indeed a confounder that limits the performance. This finding is rooted from our causal assumption: a Structural Causal Model (SCM) for the causalities among the pre-trained knowledge, sample features, and labels. Thanks to it, we propose a novel FSL paradigm: Interventional Few-Shot Learning (IFSL). Specifically, we develop three effective IFSL algorithmic implementations based on the backdoor adjustment, which is essentially a causal intervention towards the SCM of many-shot learning: the upper-bound of FSL in a causal view. It is worth noting that the contribution of IFSL is orthogonal to existing fine-tuning and meta-learning based FSL methods, hence IFSL can improve all of them, achieving a new 1-/5-shot state-of-the-art on \textit{mini}ImageNet, \textit{tiered}ImageNet, and cross-domain CUB. Code is released at https://github.com/yue-zhongqi/ifsl.