PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning to Self-Train for Semi-Supervised Few-Shot Classification

Xinzhe Li, Qianru Sun, Yaoyao Liu, Shibao Zheng, Qin Zhou, Tat-Seng Chua, Bernt Schiele

arXiv:1906.00562v2cs.CVcs.LGstat.ML

TL;DR

Few-shot classification must operate with very little labeled data, motivating methods that also use abundant unlabeled examples. LST meta-learns self-training, including pseudo-label selection and weighting, with fine-tuning after each step. Across two ImageNet benchmarks, it reports consistent improvements over state-of-the-art FSC and SSFSC methods.

  • Problem

    Few-shot classification is challenging because training data are scarce, while semi-supervised few-shot classification seeks to use much larger unlabeled sets alongside few labeled examples.

  • Method

    LST meta-learns task initialization and pseudo-label selection, weights selected pseudo-labels with SWN, and fine-tunes on labeled data after self-training steps.

  • Results

    LST achieves consistent improvements over state-of-the-art FSC and SSFSC methods on two popular few-shot benchmarks.

  • Takeaways & Limitations

    LST provides a general optimization-based approach that can use different base-learner architectures for semi-supervised few-shot classification.

Abstract

from arXiv · show

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.

1 Introduction

The paper targets semi-supervised few-shot classification, where labeled data are scarce but unlabeled data are abundant. LST integrates self-training into meta-learning while addressing noisy pseudo-labels through soft weighting and labeled-data fine-tuning.

  • Semi-supervised few-shot classification combines few labeled examples with a much larger amount of unlabeled data.
  • LST embeds self-training into the meta gradient descent paradigm for semi-supervised few-shot classification.
  • Recursive self-training can drift by adding noisy pseudo-labels, so LST meta-learns a soft weighting network and fine-tunes with labeled data after each step.
  • LST combines an inner-loop task-specific self-training procedure with outer-loop meta-learning across tasks.
  • Experiments on miniImageNet and tieredImageNet report top performance for the proposed method.

2 Related works

Related work spans supervised few-shot learning, semi-supervised learning, and semi-supervised few-shot classification. LST differs by meta-learning weights for pseudo-labeled data within a gradient-descent-based few-shot framework.

  • Few-shot classification: Few-shot classification methods include data augmentation, metric learning, memory networks, and gradient descent-based meta-learning.
  • Few-shot classification: Gradient descent-based methods learn an outer-loop initialization that adapts a base learner to a novel few-shot task in the inner loop.
  • Semi-supervised learning: Semi-supervised learning methods leverage unlabeled data through consistency regularization, moving averages, adversarial perturbations, entropy minimization, or pseudo-labeling.
  • Semi-supervised learning: Pseudo-labeling can outperform regularization-based methods when labeled data are scarce, motivating its use in LST's inner loop.
  • Semi-supervised few-shot classification: Prior semi-supervised few-shot methods use Soft k-Means or transductive propagation, whereas LST builds on classical self-training.

3 Problem definition and denotation

Each few-shot task contains a labeled support set, an unseen query set, and an unlabeled set for semi-supervised learning. Meta-learning trains across episodes so classifiers can adapt quickly to new tasks.

  • Task formulation: A conventional few-shot task contains a labeled support set S and an unseen query set Q for testing.
  • Task formulation: The unlabeled set R is added for semi-supervised learning and may contain distracting classes absent from the support set.
  • Episodic meta-learning: Meta-learning uses meta-train and meta-test phases composed of episodes rather than individual datapoints.
  • Episodic meta-learning: The meta-objective is to fast adapt a classifier to a new task rather than directly classify unseen datapoints.

4 Learning to self-train (LST)

LST combines task-level pseudo-labeling, confidence-based selection, meta-learned soft weighting, and staged self-training to exploit unlabeled data while limiting label-noise drift.

  • Pseudo-labeling: LST pseudo-labels unlabeled samples with a classifier adapted from the support set, using meta-learned parameters for task-specific prediction.The adapted classifier θT produces pseudo labels YR for the unlabeled set R.
  • Cherry-picking: LST selects the top Z confident pseudo-labeled samples per class, forming Rp with ZC samples across C classes.Hard selection precedes soft weighting to retain only the most confident predictions.
  • Cherry-picking: The soft weighting network assigns class-normalized weights to selected pseudo-labeled samples using their features and class prototype representations.Weights are normalized across C classes through a softmax layer and are intended to reduce the influence of noisy labels.
  • Self-training: Self-training first re-trains on support and weighted pseudo-labeled data, then fine-tunes on support data alone.Across T iterations, re-training occupies steps 1 through m, while fine-tuning occupies steps m + 1 through T.
  • Recursive training: After fine-tuning, LST can recursively use θT to predict improved pseudo labels on R and repeat self-training.This recursion supports repeated self-training when the unlabeled data are reused or when additional unlabeled data are available.
  • Outer-loop optimization: The outer loop meta-optimizes Φswn with an intermediate-stage query loss and Φss plus θ′ with the final-model query loss.The corresponding meta-learning rates β1 and β2 are manually set in the experiments.

5 Experiments

Experiments evaluate LST on miniImageNet and tieredImageNet under standard, ablation, distraction, recursive-training, pseudo-labeling, and generalization settings. LST achieves strong performance across benchmarks, while analyses show benefits and boundaries of recursive self-training and pseudo-label weighting.

  • Benchmark results: LST achieves the best reported accuracy on miniImageNet at 70.1% for 1-shot and 78.7% for 5-shot, improving over MTL by 8.9% and 3.2%.These results use the benchmark comparison summarized for miniImageNet.
  • Ablation studies: Hard pseudo-label selection improves miniImageNet accuracy by 3.3% for 1-shot and 1.1% for 5-shot versus no selection, while recursive repetition adds about 1% on average.Selecting more reliable samples is reported to reduce disturbance from noisy labels.
  • Ablation studies: SWN improves hard selection by 0.9% on miniImageNet for both 1-shot and 5-shot, indicating complementary effects from soft weighting and hard selection.Using SWN alone achieves results comparable to two previous SSFSC methods.
  • Recursive self-training: Recursive training improves hard,soft by 5.1% on miniImageNet 1-shot, but distractors can reduce recursive,hard performance by 0.4% and 0.5% on miniImageNet and tieredImageNet.The reported reductions might result from disturbances propagating from early recursive stages.
  • Comparisons: More than 7.4% and 4.5% improvements over related methods are reported for miniImageNet 1-shot and 5-shot using the same unlabeled-data budget.On tieredImageNet, LST surpasses TPN by 5.6% and 1.9% for 1-shot and 5-shot; it remains best with distractors.

6 Conclusions

LST is a semi-supervised few-shot classification approach that combines recursive self-training with meta-learned selection and labeling of unsupervised data. It achieves consistent improvements over state-of-the-art FSC and SSFSC methods on two popular few-shot benchmarks.

  • LST uses recursive learning-based self-training for robust inner-loop convergence and a meta-learned cherry-picking network for unsupervised data.The method selects and labels unsupervised data in the outer loop and supports optimization-based few-shot methods with different base-learner architectures.
  • LST achieves consistent improvements over state-of-the-art FSC and SSFSC methods on two popular few-shot benchmarks.

Supplementary materials

The supplementary materials extend evaluation across recursive-stage counts, tieredImageNet results, limited unlabeled data, and different backbone architectures.

  • Supplementary experiments vary recursive-training stages, report tieredImageNet results, test five unlabeled samples per class, and evaluate different backbones.

A Using different numbers of recursive stages

On miniImageNet, meta-validation evaluates classification accuracy across recursive-stage counts in 1-shot and 5-shot settings. Performance saturates after approximately six stages because all unlabeled samples have been sampled and learned.

  • Performance saturates after approximately 6 recursive stages because the model has sampled and learned all unlabeled samples.The experiments use 100 unlabeled samples per class and sample 30 for 1-shot or 50 for 5-shot at each stage.
  • Figure 5 reports meta-validation classification accuracy for different recursive-stage counts in miniImageNet 1-shot and 5-shot settings.

B Using different numbers of re-training steps

The tieredImageNet 1-shot supplementary experiment compares LST with two recursive hard-selection baselines while varying the number of re-training steps. Each curve reports performance at the final recursive stage.

  • Figure 6 compares LST with recursive, hard without meta-learning using MTL and pre-trained ResNet-12 models on tieredImageNet 1-shot.
  • Each curve reports results at the final recursive stage while the number of re-training steps varies.

C Using a small number of unlabeled samples

The experiments test LST with only five unlabeled samples per class and examine how recursive retraining steps affect 1-shot tieredImageNet accuracy.

  • C Using a small number of unlabeled samples: LST is evaluated with a limited unlabeled set of 5 samples per class, alongside Masked Soft k-Means and TPN.The comparison covers miniImageNet and tieredImageNet, with and without distracting-class unlabeled data.
  • C Using a small number of unlabeled samples: Figure 6 compares final-stage 1-shot tieredImageNet accuracy across different numbers of retraining steps for LST and two recursive hard-label baselines.Each recursive stage uses m retraining steps and 40−m fine-tuning steps in total.
  • C Using a small number of unlabeled samples: Table 5 reports classification accuracy (%) on miniImageNet and tieredImageNet under the 5-samples-per-class setting.The “w/D” condition uses unlabeled data from distracting classes excluded from the support set.

D Generalization ability

LST is tested with alternative base architectures, including MAML’s 4CONV and LEO’s ResNet-12, to assess whether its gains extend beyond the original MTL backbone.

  • D Generalization ability: LST is incorporated with MAML’s 4CONV architecture and LEO’s ResNet-12 architecture.The resulting variants are evaluated in Table 6.
Loading 1906.00562v2…