[Paper Review] A Review of Generalized Zero-Shot Learning Methods

Farhad Pourpanah, Moloud Abdar+6 more|arXiv (Cornell University)|Nov 17, 2020

Domain Adaptation and Few-Shot LearningComputer Science229 references83 citations

TL;DR

This paper provides a comprehensive survey of Generalized Zero-Shot Learning (GZSL), proposes a hierarchical categorization of methods, and discusses datasets, applications, challenges, and future directions.

ABSTRACT

Generalized zero-shot learning (GZSL) aims to train a model for classifying data samples under the condition that some output classes are unknown during supervised learning. To address this challenging task, GZSL leverages semantic information of the seen (source) and unseen (target) classes to bridge the gap between both seen and unseen classes. Since its introduction, many GZSL models have been formulated. In this review paper, we present a comprehensive review on GZSL. Firstly, we provide an overview of GZSL including the problems and challenges. Then, we introduce a hierarchical categorization for the GZSL methods and discuss the representative methods in each category. In addition, we discuss the available benchmark data sets and applications of GZSL, along with a discussion on the research gaps and directions for future investigations.

Motivation & Objective

• Provide a comprehensive review of GZSL, including problem formulation, challenges, and semantic information used.
• Introduce a hierarchical categorization of GZSL methods with representative models and applications.
• Discuss benchmark datasets and applications across domains, and identify research gaps.
• Offer directions for future research in GZSL to address biases and domain shifts.

Proposed method

• Classify GZSL approaches into embedding-based and generative-based methods and further sub-categorize embedding-based methods (graph-based, attention-based, autoencoder-based, meta-learning, compositional, bidirectional).
• Discuss problem formulation for seen/unseen classes, inductive vs. transductive settings, and calibration strategies.
• Explain the role of semantic information (attributes, word vectors) and embedding spaces (semantic, visual, latent) in building cross-domain mappings.
• Address key challenges such as hubness, projection domain shift, and bias toward seen classes, and describe mitigation strategies like calibrated stacking and novelty detectors.

Experimental results

Research questions

• RQ1 What are the core problem formulations and settings (inductive vs. transductive) in Generalized Zero-Shot Learning?
• RQ2 How can GZSL methods be systematically categorized and what are representative models in each category?
• RQ3 What datasets, benchmarks, and applications drive GZSL research, and what gaps exist in current practice?
• RQ4 What are the main challenges (hubness, domain shift, bias) and how can they be mitigated in GZSL?
• RQ5 What directions are most promising for future work in GZSL?

Key findings

• This paper provides the first in-depth, comprehensive review and analysis of GZSL methods.
• It offers a hierarchical categorization of GZSL techniques along with representative models and real-world applications.
• It discusses benchmark datasets, applications in computer vision and NLP, and identifies research gaps and future directions.
• It highlights core challenges such as hubness, projection domain shift, and bias toward seen classes, and surveys mitigation strategies like calibrated stacking and novelty detection.
• It contrasts embedding-based and generative-based approaches and details subcategories such as graph-based, meta-learning, and attention-based methods.