PaperlayerView exact revision on arXiv

Source-linked AI summary

Few-Shot Learning with Global Class Representations

Tiange Luo, Aoxue Li, Tao Xiang, Weiran Huang, Liwei Wang

arXiv:1908.05257v1cs.CV

TL;DR

Few-shot learning aims to recognize novel classes from few labeled examples, while meta-learning approaches trained mostly on source classes offer no guarantee of target-class generalization. The paper learns global class representations jointly from base and novel training samples, with episodic registration and sample synthesis. It reports effectiveness in both standard and generalized FSL, with a bigger improvement in the generalized setting.

  • Problem

    Few-shot learning seeks to classify novel classes with few labeled examples, but meta-learning models trained mostly on source classes offer no guarantee of generalization to target data.

    Source: §1 · pp. 1–2

  • Method

    The method jointly learns global representations for base and novel classes, registering support-set episodic means for query-set classification and synthesizing novel-class samples.

    Source: §1 · p. 2

  • Results

    The approach is effective in both standard and generalized FSL, with a bigger improvement under the generalized setting.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The approach extends to generalized FSL, where test labels include both base and novel classes rather than novel classes alone.

    Source: §1 · p. 2

Abstract

from arXiv · show

In this paper, we propose to tackle the challenging few-shot learning (FSL) problem by learning global class representations using both base and novel class training samples. In each training episode, an episodic class mean computed from a support set is registered with the global representation via a registration module. This produces a registered global class representation for computing the classification loss using a query set. Though following a similar episodic training pipeline as existing meta learning based approaches, our method differs significantly in that novel class training samples are involved in the training from the beginning. To compensate for the lack of novel class training samples, an effective sample synthesis strategy is developed to avoid overfitting. Importantly, by joint base-novel class training, our approach can be easily extended to a more practical yet challenging FSL setting, i.e., generalized FSL, where the label space of test data is extended to both base and novel classes. Extensive experiments show that our approach is effective for both of the two FSL settings.

1. Introduction

Few-shot learning seeks to classify novel classes from few labeled examples, but meta-learning approaches trained mostly on base-class data may not generalize to novel classes. This paper jointly learns global representations for base and novel classes, using sample synthesis and episodic registration, and reports effectiveness in both standard and generalized FSL.

1. Introduction

  • Meta-learning approaches mostly learn from source data, which offers no guarantee that their learned models generalize well to target data.FSL transfers knowledge from classes with ample training samples to novel classes with few labeled examples.Source: §1 · pp. 1–2
  • The method represents each base or novel class as a global point in embedding space, learned jointly from both groups’ training samples.The authors argue that involving novel-class data from the beginning helps ensure the model suits novel classes.Source: §1 · p. 2
  • Sample synthesis increases novel-class intra-class variation, while episodic registration uses support-set means and query-set classification loss to train the global representations.The registration module pulls a class’s global representation toward its samples and pushes other class representations away.Source: §1 · p. 2
  • Jointly training on base and novel classes is presented as enabling extension to generalized FSL, where test labels include both groups rather than novel classes alone.The standard setting tests only novel-class samples.Source: §1 · p. 2
  • The approach is effective in both standard and generalized FSL, with a bigger improvement under the generalized setting.Generalized FSL tests include both base and novel classes.Source: §1 · p. 2

2. Related Work

Prior FSL work largely uses meta-learning to transfer knowledge from auxiliary tasks, but training only on base classes leaves models vulnerable to base-class overfitting. This paper contrasts its jointly learned global representations with episodic prototypes and treats sample synthesis as a complementary strategy.

2. Related Work

  • Meta-learning FSL methods learn initialization, comparison metrics, or optimizers from base-class samples, leaving them vulnerable to overfitting to those classes.The three categories are described as learning to fine-tune, learning a good metric to compare, and learning an optimizer.Source: §2 · p. 2
  • Unlike prototypical networks, this method learns global rather than episodic class representations and jointly uses base- and novel-class training samples.Prototypical networks form class prototypes by passing a few-shot feature mean through a fully connected layer.Source: §2 · p. 2
  • Feature hallucination is described as orthogonal to this approach, and either can be used to synthesize novel-class samples against class imbalance.Source: §2 · p. 3

3. Methodology

The method jointly learns global representations for base and novel classes, using sample synthesis and episodic training to address novel-class data scarcity. Support-set class representations register to global representations, which classify query images; the same class representations support generalized FSL.

3. Methodology

  • The framework jointly optimizes global class representations, the registration module, and feature extractor using registration and query-classification losses.Each training episode constructs support-set class representations and uses registered global representations to classify query images.Source: §3 · p. 3, Figure 2 · p. 4

3.1. Registration Module

  • Registration compares training samples with global representations for all base and novel classes, learning class representations close to in-class samples and away from other classes.The registration loss uses cross-entropy to make each sample nearest to its labeled class representation in the embedding space.Source: §3.1 · p. 3

3.2. Sample Synthesis Module

  • Novel-class synthesis first augments original samples, then forms new samples from randomly selected visual features to increase intra-class variation.The first stage uses random cropping, flipping, and data hallucination; the second samples from the subspace spanned by selected feature vectors.Source: §3.2 · p. 3

3.3. Few Shot Learning By Registration

  • Training episodes sample classes and support/query examples; when novel-class labels are insufficient, synthesis expands them to the required episode size.Support-set features form episodic class representations: base-class features are averaged, while novel-class representations use synthesized samples. These register global representations, and query classification and registration losses jointly update model components.Source: §3.3 · pp. 4–5

3.4. Extension to Generalized FSL

  • Generalized FSL includes test data from both base and novel classes and predicts labels across all classes in Ctotal.The registration module compares each test sample against global representations for both class groups.Source: §3.4 · pp. 5–6

4. Experiments and Discussion

The experiments evaluate the approach in standard FSL, generalized FSL, and ablation settings.

4. Experiments and Discussion

  • The standard FSL experiments restrict each test episode’s label space to a few novel classes.This is one of three experimental groups.Source: §4 · p. 6
  • The generalized FSL experiments extend the test label space to both base and novel classes.This is one of three experimental groups.Source: §4 · p. 6
  • The experiments also include an ablation study.The passage identifies this as the third experimental group.Source: §4 · p. 6

4.1. Standard Few-Shot Learning

The standard few-shot evaluation uses Omniglot and miniImageNet benchmarks and reports state-of-the-art Omniglot performance, with strong results on miniImageNet.

4.1.1 Datasets and Settings

  • Evaluation covers Omniglot and miniImageNet, using 1,200 Omniglot training characters and a 64/16/20-class miniImageNet train/validation/test split.The reported miniImageNet model uses 64 base classes and 20 novel test classes.Source: §4.1.1 · p. 6

4.1.2 Implementation Details

  • Training initializes the feature extractor on base-class classification, then jointly trains it with global representations, the data hallucinator, and registration module.The registration module starts from random Gaussian initialization.Source: §4.1.2 · p. 6

4.1.3 Results on Omniglot

  • Omniglot evaluation includes 5-way and 20-way classification with both 1-shot and 5-shot tasks, scored over 1,000 test episodes.Each query set contains five images per class.Source: §4.1.3 · pp. 6–7
  • The approach achieves a new state-of-the-art result on Omniglot.The authors attribute this to their global class representation learning strategy.Source: §4.1.3 · p. 7

4.1.4 Results on miniImageNet

  • On miniImageNet, the approach significantly outperforms alternatives in 5-way 5-shot and ties for best in 5-way 1-shot.The registration module also reaches 100% registration accuracy on test data.Source: §4.1.4 · p. 7, Table 2 · p. 7

4.2. Generalized Few-Shot Learning

Generalized FSL tests both base and novel classes in a joint label space. The approach leads all three reported metrics, with larger margins than in standard FSL.

4.2.1 Dataset and Settings

  • The miniImageNet experiment uses a new sample split and evaluates 5-way 5-shot learning with 100 test images per class.Training draws 500 images per base class and a few samples per novel class; test labels are predicted over 100 classes.Source: §4.2.1 · pp. 7–8
  • The generalized-FSL test label space includes both base and novel classes, and evaluation measures overall, base-origin, and novel-origin classification accuracy.The metrics are acca, accb, and accun, respectively.Source: §4.2.1 · p. 7
  • Figure 3 defines B as within-class feature averaging and labels the method components S1, S2, and R.S1 and S2 are the two steps in Section 3.2; R is the registration module from Section 3.1.Source: Figure 3 · p. 8

4.2.2 Results

  • The approach achieves the best results on all three generalized-FSL metrics, with bigger margins than under standard FSL.The authors attribute its advantage over PN and RN to learning global rather than episodic class representations.Source: §4.2.2 · p. 8

4.3. Ablation Study

The component analysis finds that sample synthesis or registration alone is insufficient, while combining them improves performance. The new-class experiment reports that the method beats Prototypical Net even with seen-class representations fixed.

4.3.1 Components Analysis

  • The ablation compares six models built from class averaging (B), two synthesis steps (S1 and S2), and registration (R), with B+S1+S2+R as the full model.The variants use the same FSL framework and differ in how they learn global class representations.Source: §4.3.1 · p. 8
  • Sample synthesis or registration alone cannot learn class global representations well, but using both significantly improves performance.The comparisons include B+S1, B+S1+S2, and B+R versus B, and B+S1+R and B+S1+S2+R versus B.Source: §4.3.1 · p. 8

4.3.2 New Novel Classes

  • With the 100 seen-class representations fixed, the method beats Prototypical Net by a large margin when adding 20 new classes.The generalized FSL test covers the joint label space of all 120 classes; each new class has 5 training and 100 test samples.Source: §4.3.2 · p. 9, Table 4 · p. 9

5. Conclusion

The conclusion presents global class representations learned from base and novel samples as the paper’s approach to few-shot learning. It reports effectiveness in standard and generalized FSL.

5. Conclusion

  • The approach learns global class representations using both base- and novel-class training samples.In each episode, a support-set class mean is registered with the global representation to produce a representation used for query-set classification loss.Source: §5 · p. 9
  • The approach extends to generalized FSL, where the test label space includes both base and novel classes.Source: §5 · p. 9
  • The paper reports effectiveness in both standard and generalized FSL.Source: §5 · p. 9

Related papers

Loading 1908.05257v1…