Core Concepts
MetaCoCo, a large-scale benchmark with spurious-correlation shifts collected from real-world scenarios, is presented to facilitate the development of models robust to spurious-correlation shifts in few-shot classification.
Abstract
The paper presents MetaCoCo, a new few-shot classification benchmark that focuses on the problem of spurious-correlation shifts. Unlike existing few-shot classification benchmarks that mainly address cross-domain shifts, MetaCoCo introduces spurious-correlation shifts by incorporating contextual information into the images.
The key highlights of the paper are:
Motivation and Problem Definition:
Out-of-distribution (OOD) problems in few-shot classification can be categorized into cross-domain shifts and spurious-correlation shifts.
Existing benchmarks mainly focus on cross-domain shifts, while spurious-correlation shifts remain understudied due to lack of corresponding evaluation benchmarks.
MetaCoCo Benchmark:
MetaCoCo is a large-scale benchmark with 175,637 images, 155 contexts, and 100 classes, designed to reflect spurious-correlation shifts in real-world scenarios.
Each class (concept) is associated with various contexts, introducing spurious correlations between concepts and contexts.
The benchmark is divided into training, validation, and testing sets to evaluate the performance of few-shot classification methods under spurious-correlation shifts.
Evaluation Metrics:
A novel metric is proposed using CLIP, a pre-trained vision-language model, to quantify the extent of spurious correlations in MetaCoCo and other few-shot classification benchmarks.
Experimental Evaluation:
Extensive experiments are conducted on MetaCoCo to evaluate the state-of-the-art methods in few-shot classification, cross-domain shifts, and self-supervised learning.
The results show that the performance of existing methods degrades significantly in the presence of spurious-correlation shifts, highlighting the importance of addressing this problem.
The proposed MetaCoCo benchmark aims to facilitate future research on spurious-correlation shifts in few-shot classification, which is a crucial and understudied problem in real-world applications.
Stats
"The performance of all methods decreases compared with existing FSC benchmarks, which demonstrates that these methods are insufficient in solving the spurious-correlation-shift problem."
"The advantages of cross-domain few-shot classification methods disappear on MetaCoCo, resulting in weaker performance, even worse than non-cross-domain FSC methods."
Quotes
"The shifts in cross-domain benchmarks are caused by varying distributions between various datasets. Instead, the shifts in MetaCoCo are caused by varying both concepts and contexts."
"Once images do not match the contexts, the performance will deteriorate."
"It remains unclear what is the best learning strategy for avoiding the effect of spurious-correlation contexts and the most appropriate episodic sample."