[Paper Review] Assume, Augment and Learn: Unsupervised Few-Shot Meta-Learning via Random Labels and Data Augmentation

Antreas Antoniou, Amos Storkey|arXiv (Cornell University)|Feb 26, 2019

Domain Adaptation and Few-Shot LearningComputer Science22 references36 citations

TL;DR

Introduces AAL, an unsupervised method to generate few-shot tasks from unlabeled data by assigning random labels to a subset and creating a target set through data augmentation; usable with standard meta-learning frameworks to perform well on few-shot tasks without fine-tuning.

ABSTRACT

The field of few-shot learning has been laboriously explored in the supervised setting, where per-class labels are available. On the other hand, the unsupervised few-shot learning setting, where no labels of any kind are required, has seen little investigation. We propose a method, named Assume, Augment and Learn or AAL, for generating few-shot tasks using unlabeled data. We randomly label a random subset of images from an unlabeled dataset to generate a support set. Then by applying data augmentation on the support set's images, and reusing the support set's labels, we obtain a target set. The resulting few-shot tasks can be used to train any standard meta-learning framework. Once trained, such a model, can be directly applied on small real-labeled datasets without any changes or fine-tuning required. In our experiments, the learned models achieve good generalization performance in a variety of established few-shot learning tasks on Omniglot and Mini-Imagenet.

Motivation & Objective

• Motivate unsupervised few-shot learning where no labels are available.
• Propose a task-generation strategy that creates support and target sets from unlabeled data.
• Enable existing meta-learning frameworks to train on these unsupervised tasks without fine-tuning.
• Evaluate the approach on standard few-shot benchmarks to assess generalization.
• Investigate data augmentation strategies to maximize transfer to real labeled data.

Proposed method

• Create support sets by randomly assigning labels to randomly sampled unlabeled data points.
• Generate corresponding target sets by applying data augmentation to the support set and keeping labels fixed.
• Train standard meta-learning models (MAML and Prototypical Networks) on the generated unsupervised tasks.
• Evaluate trained models on real labeled few-shot tasks without fine-tuning.
• Explore and ablate various data augmentation techniques to determine effective augmentations for target sets.
• Apply the approach within the MAML++ framework to improve stability and generalization.

Experimental results

Research questions

• RQ1 Can unsupervised task generation from unlabeled data enable effective few-shot meta-learning without any labels?
• RQ2 Which data augmentation strategies best transfer learned representations to real labeled few-shot tasks?
• RQ3 How does the unsupervised AAL approach perform when combined with MAML and Prototypical Networks?
• RQ4 What is the impact of augmentations on generalization to Omniglot and Mini-Imagenet benchmarks?

Key findings

• The AAL approach enables few-shot learning models to generalize to real labeled tasks without fine-tuning.
• Unsupervised tasks generated from unlabeled data can be used with standard meta-learning frameworks (MAML and Prototypical Networks).
• Data augmentation plays a crucial role in creating meaningful target sets and can significantly affect generalization performance.
• In experiments, AAL with MAML++ and certain augmentations achieves competitive results on Omniglot and Mini-Imagenet tasks compared to supervised baselines and concurrent unsupervised methods.
• The method demonstrates better performance than at least one concurrent unsupervised baseline on Omniglot tasks, with potential trade-offs when leveraging auto-augmented augmentations from large labeled datasets.