An Iterative Co-Training Transductive Framework for Zero Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Transductive learning in zero-shot learning (ZSL) generally outperforms inductive learning by utilizing unseen-class samples.
• Generating accurate pseudo-labels for unseen classes and managing their inherent noise are key challenges in transductive ZSL.

Purpose of the Study:

• To propose an iterative co-training framework that enhances transductive zero-shot learning performance.
• To adapt the co-training framework for generalized zero-shot learning (GZSL) by addressing the bias problem.

Main Methods:

• An iterative co-training framework with two distinct ZSL models and an exchanging module was developed.
• Models iteratively predict pseudo-labels for unseen-class samples, which are then exchanged and added to training sets.
• A semantic-guided Out-of-Distribution (OOD) detector was integrated for GZSL to identify likely unseen-class samples.

Main Results:

• The proposed framework significantly boosts ZSL performance by exploiting model complementarity.
• The framework demonstrated superior performance in generalized ZSL (GZSL) by mitigating bias.
• Extensive experiments on three benchmarks showed outperformance against 31 state-of-the-art methods.

Conclusions:

• The iterative co-training framework effectively improves transductive ZSL by leveraging diverse model predictions.
• The approach offers a robust solution for GZSL by incorporating an OOD detector to handle bias.
• The proposed methods represent a significant advancement in ZSL and GZSL research.