Rebalanced Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) identifies unseen classes without training samples.
• Current ZSL methods often suffer from imbalanced semantic predictions, performing well for some semantics but poorly for others.
• Traditional imbalanced learning approaches are insufficient for ZSL's unique challenges.

Purpose of the Study:

• To introduce an imbalanced learning framework tailored for ZSL.
• To address the unique challenges of imbalanced ZSL, including label value correlation and diverse error distributions.
• To improve the accuracy and robustness of ZSL models.

Main Methods:

• Formalized ZSL as an imbalanced regression problem to understand semantic label influence.
• Proposed a novel re-weighted loss function: Re-balanced Mean-Squared Error (ReMSE).
• ReMSE tracks error distribution mean and variance for balanced class learning.

Main Results:

• Empirical evidence demonstrates how semantic labels cause imbalanced predictions in ZSL.
• ReMSE effectively alleviates semantic prediction imbalance across classes.
• Extensive experiments show the proposed method outperforms state-of-the-art ZSL techniques.

Conclusions:

• The proposed imbalanced learning framework and ReMSE loss offer a theoretically sound solution for ZSL.
• ReMSE significantly enhances ZSL performance by addressing inherent prediction imbalances.
• This work advances the field of ZSL by providing a more robust and accurate approach.