Zero-shot learning via visual-semantic aligned autoencoder.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) aims to recognize unseen classes using knowledge from seen classes and semantic features.
• Existing generative methods for ZSL often introduce bias, as unseen samples are generated after model training on seen data.

Purpose of the Study:

• To develop a novel method for generating unseen class samples in zero-shot learning that mitigates bias.
• To improve the accuracy and reliability of zero-shot recognition by addressing limitations of current generative approaches.

Main Methods:

• Utilized an autoencoder model to generate unseen class samples.
• Integrated semantic features of unseen classes with newly generated sample features.
• Constructed a specialized loss function incorporating these combined features.

Main Results:

• The proposed autoencoder-based method demonstrated improved generation of unseen class samples.
• Experimental validation on three datasets confirmed the effectiveness of the approach.
• The method showed reduced bias compared to traditional generative models.

Conclusions:

• The autoencoder approach offers a promising solution for generating unbiased unseen class samples in zero-shot learning.
• This technique enhances the performance of zero-shot recognition by providing more accurate representations of novel classes.