A Discriminative Cross-Aligned Variational Autoencoder for Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) aims to classify unseen data by leveraging relationships between visual and semantic features.
• Traditional ZSL methods often struggle with hubness and domain bias, limiting performance, especially in generalized ZSL (GZSL).

Purpose of the Study:

• To develop a novel model, the Discriminative Cross-Aligned Variational Autoencoder (DCA-VAE), to overcome limitations in traditional ZSL and GZSL.
• To improve the classification of unseen samples by effectively aligning multimodal data structures.

Main Methods:

• The proposed DCA-VAE utilizes a modified variational autoencoder (VAE) with a discriminative cosine metric.
• It transforms visual and semantic features into latent features by capturing principal discriminative information.
• This approach constructs latent features rich in discriminative multimodal information for unseen samples.

Main Results:

• The DCA-VAE model was validated on six benchmark datasets, including ImageNet.
• Experimental results demonstrate the superiority of DCA-VAE over existing embedding and generative ZSL models.
• The model shows improved performance on both standard ZSL and the more challenging GZSL tasks.

Conclusions:

• The DCA-VAE effectively addresses hubness and domain bias issues in ZSL.
• The proposed method enhances the classification of unseen samples by learning discriminative latent representations.
• DCA-VAE offers a superior approach for both ZSL and GZSL tasks compared to existing methods.