An Interpretable Deep Architecture for Similarity Learning Built Upon Hierarchical Concepts

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Complex mathematical models like deep neural networks enhance learning accuracy but reduce interpretability.
• Understanding the internal workings of complex models is crucial for trust and debugging, especially in image retrieval tasks.

Purpose of the Study:

• To propose an effective Similarity Neural Network (SNN) that achieves robust image retrieval performance.
• To address the challenge of reduced post-hoc model interpretability in complex learning models.
• To provide methods for understanding and visualizing the learned features within the SNN.

Main Methods:

• Designed a Similarity Neural Network (SNN) by integrating neuron architecture with a concept tree organization.
• Formulated neuron operations to facilitate the passing of similarity information between concepts.
• Developed techniques for visualizing and interpreting the learned representations within the SNN.

Main Results:

• The proposed SNN demonstrated superior performance compared to state-of-the-art approaches in image retrieval tasks.
• Neuron visualization results provided insights into the network's learning process and feature representation.
• The SNN achieved a favorable balance between retrieval accuracy and post-hoc interpretability.

Conclusions:

• The developed SNN offers an effective solution for image retrieval, enhancing both performance and interpretability.
• The proposed visualization methods aid in understanding complex deep learning models.
• This work contributes to the development of more transparent and reliable AI systems for similarity learning.