Combining Deep Learning and Case-Based Reasoning for Robust, Accurate, Explainable Classification
Abstract
This proposal will investigate a path towards reliable, explainable, accurate machine learning by uniting deep learning with case-based reasoning (CBR), which are models that explicitly encode domain knowledge including common sense and human expertise. CBR makes decisions by consulting key training examples seen in the past through domain-specific similarity rules, adapting them to the target problem through adaptation rules, testing this solution, and using feedback to decide to add the example to the knowledge base. This approach has many advantages over deep learning, because it: (1) allows human expertise to be directly encoded by asking them to write cases and adaptation/similarity rules, (2) can reason from limited training data given appropriate adaptation and similarity rules, (3) improves over time in an inertia-free way without requiring retraining, (4) can explain its answers by citing the cases and adaptation rules that it used, and (5) mimics the prototype theory that humans are thought to use. Despite these advantages and numerous practical successes of CBR, deep learning has become much more popular because it delivers better accuracy on many practical problems when large datasets are available, can better discover statistical patterns that may not be apparent to a human, and high-quality, easy-to-use software frameworks are available.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 30, 2019
- Source ID
- N000141912655
Entities
People
- David Crandall
Organizations
- Indiana University
- Office of Naval Research
- United States Navy