Meta-Learning for Generalized Zero-Shot Learning

Vinay Kumar Verma, Dhanajit Brahma, Piyush Rai

Abstract

Learning to classify unseen class samples at test time is popularly referred to as zero-shot learning (ZSL). If test samples can be from training (seen) as well as unseen classes, it is a more challenging problem due to the existence of strong bias towards seen classes. This problem is generally known as generalized zero-shot learning (GZSL). Thanks to the recent advances in generative models such as VAEs and GANs, sample synthesis based approaches have gained considerable attention for solving this problem. These approaches are able to handle the problem of class bias by synthesizing unseen class samples. However, these ZSL/GZSL models suffer due to the following key limitations: (i) Their training stage learns a class-conditioned generator using only seen class data and the training stage does not explicitly learn to generate the unseen class samples; (ii) They do not learn a generic optimal parameter which can easily generalize for both seen and unseen class generation; and (iii) If we only have access to a very few samples per seen class, these models tend to perform poorly. In this paper, we propose a meta-learning based generative model that naturally handles these limitations. The proposed model is based on integrating model-agnostic meta learning with a Wasserstein GAN (WGAN) to handle (i) and (iii), and uses a novel task distribution to handle (ii). Our proposed model yields significant improvements on standard ZSL as well as more challenging GZSL setting. In ZSL setting, our model yields 4.5%, 6.0%, 9.8%, and 27.9% relative improvements over the current state-of-the-art on CUB, AWA1, AWA2, and aPY datasets, respectively.

Introduction

In Zero-Shot Learning (ZSL), a model must classify objects from classes it has never seen during training, relying instead on auxiliary information like text descriptions or attribute vectors. Generalized Zero-Shot Learning (GZSL) is even more challenging: at test time, the model must distinguish between both seen and unseen classes.

Current generative approaches for GZSL—which synthesize artificial examples of unseen classes to train a classifier—suffer from three main flaws: a lack of explicit generalization to unseen classes during training, high sensitivity to parameter initialization, and poor performance in low-data regimes where labeled examples are scarce.

Methodology: The ZSML Framework

The authors propose ZSML (Zero-Shot Meta-Learning), which integrates a conditional Wasserstein GAN (WGAN) into a meta-learning framework.

Architecture

The system consists of three modules: a Generator (G) that creates features from noise and class attributes, a Discriminator (D) that distinguishes real from synthetic features, and a Classifier (C) that ensures the generated features belong to the intended class. As shown in Figure 1, meta-learners are used to optimize the parameters of these modules to find an initialization that generalizes across the seen-unseen divide.

The ZSML architecture integrates a Generator, Discriminator, and Classifier, each optimized by a Meta-Learner to find generalizable parameters.

Disjoint Task Distribution

The core innovation of ZSML is its novel task distribution. Unlike standard meta-learning where training and validation sets share the same classes, ZSML ensures that the classes in the meta-training set and meta-validation set are disjoint. This forces the generator to learn how to produce features for classes it hasn't seen yet within every single training step, mimicking the actual test scenario.

Experimental Results

The authors evaluated ZSML on five standard ZSL benchmarks: CUB, AWA1, AWA2, aPY, and SUN.

Performance and Robustness

ZSML achieved state-of-the-art (SOTA) results across the board. In standard ZSL, the model showed relative improvements of 4.5% to 27.9% over previous methods. In the more difficult GZSL setting, ZSML significantly improved the Harmonic Mean (balancing seen and unseen class accuracy). For example, on AWA2, it achieved a Harmonic Mean of 65.8, outperforming previous models like f-CLSWGAN (59.6).

Comparative performance of ZSML against baselines (top) and its robustness in few-shot scenarios (bottom).

As illustrated in Figure 2, ZSML remains highly effective even in the "low-data" regime. When only 5 or 10 samples per seen class are available, it outperforms existing methods by a wide margin, proving that the meta-learning approach makes the model much more data-efficient.

Conclusion

ZSML addresses the core limitations of generative GZSL by treating unseen class generation as a skill to be mastered during training. While performance on the high-cardinality SUN dataset suggests potential issues with mode collapse, the framework's overall success across diverse benchmarks demonstrates the power of meta-learning for zero-shot tasks. Future work will explore different generative architectures, such as VAEs, to enhance stability on massive, fine-grained datasets.

Summary updated

Abstract

Learning to classify unseen class samples at test time is popularly referred to as zero-shot learning (ZSL). If test samples can be from training (seen) as well as unseen classes, it is a more challenging problem due to the existence of strong bias towards seen classes. This problem is generally known as generalized zero-shot learning (GZSL). Thanks to the recent advances in generative models such as VAEs and GANs, sample synthesis based approaches have gained considerable attention for solving this problem. These approaches are able to handle the problem of class bias by synthesizing unseen class samples. However, these ZSL/GZSL models suffer due to the following key limitations: (i) Their training stage learns a class-conditioned generator using only seen class data and the training stage does not explicitly learn to generate the unseen class samples; (ii) They do not learn a generic optimal parameter which can easily generalize for both seen and unseen class generation; and (iii) If we only have access to a very few samples per seen class, these models tend to perform poorly. In this paper, we propose a meta-learning based generative model that naturally handles these limitations. The proposed model is based on integrating model-agnostic meta learning with a Wasserstein GAN (WGAN) to handle (i) and (iii), and uses a novel task distribution to handle (ii). Our proposed model yields significant improvements on standard ZSL as well as more challenging GZSL setting. In ZSL setting, our model yields 4.5%, 6.0%, 9.8%, and 27.9% relative improvements over the current state-of-the-art on CUB, AWA1, AWA2, and aPY datasets, respectively.
BibTeX
@inproceedings{verma2020meta,
  author = {Vinay Kumar Verma and Dhanajit Brahma and Piyush Rai},
  title = {{Meta-Learning for Generalized Zero-Shot Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  url = {https://openreview.net/forum?id=4PK8vow_fw},
}
Figure