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 several key limitations: they learn generators using only seen data without explicit practice on unseen generation, they lack generic optimal parameters for both seen and unseen classes, and they perform poorly with limited training samples. In this paper, we propose a meta-learning based generative model, ZSML, that naturally handles these limitations by integrating model-agnostic meta learning with a Wasserstein GAN (WGAN) and utilizing a novel disjoint task distribution.
Introduction
Zero-Shot Learning (ZSL) aims to recognize objects from categories not seen during training by leveraging auxiliary information like semantic attributes. Generalized Zero-Shot Learning (GZSL) extends this by requiring the model to classify both seen and unseen classes at test time. While generative models like GANs have improved GZSL by synthesizing features for unseen classes, they often suffer from a "generalization gap" because they are trained exclusively on seen classes. This leads to a strong bias where unseen samples are misclassified as seen ones. Furthermore, these models typically require large amounts of data per seen class, failing in low-data regimes.
Methodology
The proposed Zero-Shot Meta-Learning (ZSML) framework addresses these issues by framing feature generation as a meta-learning problem.
Meta-Learning with WGAN
ZSML integrates Model-Agnostic Meta-Learning (MAML) with a Wasserstein GAN (WGAN). Instead of learning parameters for a specific set of seen classes, the model learns a set of "meta-parameters" that are optimized for fast adaptation to new, unseen class distributions. The architecture consists of three primary modules: a Generator ($G$), a Discriminator ($D$), and a Classifier ($C$). All three modules are optimized using the meta-learning loop to ensure they can generalize effectively.
Disjoint Task Distribution
A critical innovation in ZSML is the use of a disjoint task distribution during meta-training. In each training episode, the classes used for the "meta-train" set and the "meta-validation" set are strictly disjoint. This forces the generator to learn the underlying mapping from semantic attributes to visual features in a way that generalizes to entirely new categories, effectively "rehearsing" the zero-shot scenario during training.
Experimental Results
The model was evaluated on five standard benchmarks: CUB (fine-grained birds), AWA1 & AWA2 (animals), aPY (diverse objects), and SUN (scenes). Following standard protocol, ResNet-101 features were used as inputs.
As shown in the results table, ZSML achieves state-of-the-art performance across nearly all datasets. Notably, it provides a massive 27.9% relative improvement on the aPY dataset. In the GZSL setting, ZSML consistently yields higher Harmonic Mean scores, indicating a significant reduction in seen-class bias.
Robustness to Data Scarcity
One of the most impressive features of ZSML is its performance in "Few-Shot ZSL." While traditional generative models experience sharp performance drops when training samples per seen class are limited (e.g., to 10 samples), ZSML maintains high accuracy. This robustness stems from the meta-learning objective, which is inherently designed for effective learning from few examples.
Conclusion and Discussion
ZSML successfully bridges the gap between seen and unseen class distributions by incorporating meta-learning into the generative process. By mimicking the zero-shot test scenario through disjoint task distributions, the model learns to synthesize high-quality features for unseen classes. While ZSML shows remarkable gains, challenges remain on datasets with very high class counts like SUN, where GAN-based models can struggle with mode collapse. Future work may involve exploring more stable generative architectures to further enhance performance.
Summary updated