Building and Interpreting Deep Similarity Models.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Many machine learning algorithms rely on distance or similarity measures.
• Verifying that these similarities reflect meaningful data patterns is crucial before model training.
• Interpretability of similarity models is essential for trust and validation.

Purpose of the Study:

• To develop a method for making deep similarity models interpretable.
• To provide systematic explanations for the similarity outputs of trained models.
• To enhance the verifiability of similarity-based machine learning approaches.

Main Methods:

• Introduction of BiLRP, a scalable and theoretically grounded method.
• BiLRP decomposes the output of deep similarity models on input feature pairs.
• The method leverages Layer-wise Relevance Propagation (LRP) explanations for scalability with non-linear models.

Main Results:

• BiLRP successfully and robustly explains complex similarity models, including those using VGG-16 deep neural network features.
• The method was applied to assess similarity between historical documents, such as astronomical tables, in digital humanities.
• BiLRP provided significant insights and verifiability for a problem-specific similarity model in digital humanities.

Conclusions:

• BiLRP offers a powerful tool for interpreting deep similarity models across various domains.
• The method enhances the trustworthiness and applicability of similarity-based machine learning.
• BiLRP demonstrates utility in both general machine learning and specialized fields like digital humanities.