LAVA: Label-efficient Visual Learning and Adaptation

TL;DR

LAVA is a novel method for multi-domain visual transfer learning that effectively utilizes limited labeled and unlabeled data through self-supervision and robust pseudo-labeling, achieving state-of-the-art results.

Contribution

LAVA introduces a new approach combining self-supervised representations and multi-crop pseudo-labeling for efficient transfer learning with limited data.

Findings

Achieves state-of-the-art on ImageNet semi-supervised protocol

Outperforms existing methods on 7 out of 10 Meta-dataset tasks

Effectively handles class and domain shifts with limited labeled data

Abstract

We present LAVA, a simple yet effective method for multi-domain visual transfer learning with limited data. LAVA builds on a few recent innovations to enable adapting to partially labelled datasets with class and domain shifts. First, LAVA learns self-supervised visual representations on the source dataset and ground them using class label semantics to overcome transfer collapse problems associated with supervised pretraining. Secondly, LAVA maximises the gains from unlabelled target data via a novel method which uses multi-crop augmentations to obtain highly robust pseudo-labels. By combining these ingredients, LAVA achieves a new state-of-the-art on ImageNet semi-supervised protocol, as well as on 7 out of 10 datasets in multi-domain few-shot learning on the Meta-dataset. Code and models are made available.