Towards Improved and Interpretable Deep Metric Learning via Attentive Grouping

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Grouping is a common technique in deep metric learning to generate diverse features.
• Existing methods lack interpretability and flexibility within various metric learning frameworks.

Purpose of the Study:

• To propose an improved and interpretable grouping method for deep metric learning.
• To enhance feature diversity, performance, and model interpretability.

Main Methods:

• An attention mechanism with a learnable query for each group is employed.
• The method incorporates a diversity loss for capturing group-specific information.
• The approach is formally shown to be invariant to spatial permutations, inducing translational invariance in CNNs.

Main Results:

• The proposed method consistently and significantly outperforms prior methods across various datasets, metrics, base models, and loss functions.
• Interpretation results demonstrate the learning of diverse and stable semantic features across groups.
• Attention scores provide clear interpretability by indicating feature importance at spatial positions.

Conclusions:

• The novel grouping method offers enhanced performance and interpretability in deep metric learning.
• It provides a flexible module for integration into existing convolutional neural networks.
• The method facilitates the learning of semantically meaningful and stable features.