Advances in MetaDL: AAAI 2021 challenge and workshop

TL;DR

This paper reports on the 2021 MetaDL challenge and workshop, focusing on few-shot image classification, highlighting effective methods that leverage pre-trained CNN backbones and fine-tuning under computational constraints.

Contribution

It introduces the design and results of a MetaDL challenge, emphasizing practical solution strategies for few-shot learning with limited computational resources.

Findings

Winning methods used classifiers on CNN features with fine-tuning.

Solutions relied on existing architectures and pre-trained networks.

The challenge demonstrated effective approaches under tight computational constraints.

Abstract

To stimulate advances in metalearning using deep learning techniques (MetaDL), we organized in 2021 a challenge and an associated workshop. This paper presents the design of the challenge and its results, and summarizes presentations made at the workshop. The challenge focused on few-shot learning classification tasks of small images. Participants' code submissions were run in a uniform manner, under tight computational constraints. This put pressure on solution designs to use existing architecture backbones and/or pre-trained networks. Winning methods featured various classifiers trained on top of the second last layer of popular CNN backbones, fined-tuned on the meta-training data (not necessarily in an episodic manner), then trained on the labeled support and tested on the unlabeled query sets of the meta-test data.