[Paper Review] Learning to Self-Train for Semi-Supervised Few-Shot Classification

Xinzhe Li, Qianru Sun+5 more|arXiv (Cornell University)|Jun 3, 2019

Domain Adaptation and Few-Shot Learning31 references84 citations

TL;DR

The paper introduces Learning to Self-Train (LST), a semi-supervised meta-learning approach that self-labels unlabeled data in few-shot tasks and uses a soft weighting network to select and weight pseudo-labels for robust self-training.

ABSTRACT

Few-shot classification (FSC) is challenging due to the scarcity of labeled training data (e.g. only one labeled data point per class). Meta-learning has shown to achieve promising results by learning to initialize a classification model for FSC. In this paper we propose a novel semi-supervised meta-learning method called learning to self-train (LST) that leverages unlabeled data and specifically meta-learns how to cherry-pick and label such unsupervised data to further improve performance. To this end, we train the LST model through a large number of semi-supervised few-shot tasks. On each task, we train a few-shot model to predict pseudo labels for unlabeled data, and then iterate the self-training steps on labeled and pseudo-labeled data with each step followed by fine-tuning. We additionally learn a soft weighting network (SWN) to optimize the self-training weights of pseudo labels so that better ones can contribute more to gradient descent optimization. We evaluate our LST method on two ImageNet benchmarks for semi-supervised few-shot classification and achieve large improvements over the state-of-the-art method. Code is at https://github.com/xinzheli1217/learning-to-self-train.

Motivation & Objective

• Motivate semi-supervised learning for few-shot classification (SSFSC) to utilize unlabeled data when labeled data is scarce.
• Integrate self-training into meta-learning to rapidly adapt to new few-shot tasks.
• Mitigate label-noise drift by learning soft weights for pseudo-labels and by re-training with labeled data after each step.
• Demonstrate that LST yields state-of-the-art results on miniImageNet and tieredImageNet benchmarks.

Proposed method

• Employ an inner-loop self-training process for a task starting from a meta-learned initialization.
• Generate pseudo labels for unlabeled data using a task-specific classifier trained on the support set.
• Hard-select a subset of pseudo-labeled samples and assign soft weights via a meta-learned Soft Weighting Network (SWN).
• Re-train the model on weighted pseudo-labeled data plus labeled support data, followed by a fine-tuning step on labeled data only.
• Meta-optimize SWN and base-learner initializations through outer-loop validation losses to improve future self-training steps.

Experimental results

Research questions

• RQ1 Can self-training with pseudo-labels improve performance in semi-supervised few-shot classification?
• RQ2 How can we mitigate drift from noisy pseudo-labels in iterative self-training within a meta-learning framework?
• RQ3 Does meta-learning a weighting mechanism for pseudo-labels (SWN) improve cherry-picking and overall adaptation to new tasks?
• RQ4 What are the benefits and limitations of recursive/self-trained steps under distractor unlabeled data?

Key findings

• LST achieves top results on miniImageNet (1-shot: 70.1%, 5-shot: 78.7%) and tieredImageNet (1-shot: 77.7%, 5-shot: 85.2%).
• Compared with the meta-learning baseline (MTL), LST improves miniImageNet by 8.9% (1-shot) and 3.2% (5-shot).
• On tieredImageNet, LST improves by 12.1% (1-shot) and 6.6% (5-shot) over the baseline.
• Hard selection of pseudo-labels generally boosts performance, and combining hard selection with soft weighting (SWN) yields further gains.
• Recursive self-training can improve performance, especially when combined with SWN and hard selection, though distractors can hurt gains in some settings.