Visual-guided attentive attributes embedding for zero-shot learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Zero-shot learning (ZSL) classifies unseen classes using seen class data and external knowledge, often attributes.
• Existing ZSL methods treat attributes equally, overlooking their varying importance during training.

Purpose of the Study:

• To propose a novel encoder-decoder framework with an attention mechanism for adaptive attribute exploitation in ZSL.
• To improve the discriminative power of attribute-based ZSL by emphasizing important attributes.

Main Methods:

• Developed an encoder-decoder framework mapping visual features to a semantic space.
• Incorporated an attention mechanism to assign weights based on attribute discriminability.
• Simultaneously decoded attentive attributes and class prototypes to the visual space to mitigate the hubness problem.

Main Results:

• The proposed model significantly outperforms state-of-the-art methods on ZSL tasks.
• Achieved competitive results for the generalized ZSL task.
• Demonstrated the effectiveness of adaptive attribute weighting in ZSL.

Conclusions:

• The attention-based encoder-decoder framework offers a significant advancement in zero-shot learning.
• Adaptive emphasis on discriminative attributes enhances model performance and generalization.
• The method effectively addresses challenges like the hubness problem in ZSL.