Zero-Shot Learning With Transferred Samples.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) enables model training for unseen classes using knowledge from labeled source classes.
• Conventional ZSL methods use a two-step process, involving an intermediate space projection, which can lead to information loss and reduced performance.

Purpose of the Study:

• To propose a novel one-step recognition framework for zero-shot learning.
• To overcome the limitations of conventional two-step ZSL approaches by performing recognition directly in the original feature space.

Main Methods:

• Developed a one-step recognition framework that utilizes directly trained classifiers.
• Implemented a sample transfer strategy from source to target classes, assigning pseudo-labels and selecting samples based on transferability and diversity.
• Modified the standard support vector machine (SVM) formulation to handle unreliable pseudo-labeled samples by recognizing and suppressing them during training.

Main Results:

• The proposed framework demonstrates superior performance compared to state-of-the-art approaches across four benchmark datasets.
• The one-step recognition strategy avoids information loss associated with intermediate transformations.
• The method effectively handles unreliable pseudo-labels, enhancing model robustness.

Conclusions:

• The novel one-step ZSL framework offers a more effective approach to recognition for unseen classes.
• The proposed techniques for sample transfer and modified SVM training contribute to improved ZSL performance.
• The framework's generality allows for extensions to various common ZSL settings.