Knowledge Extraction and Injection Via Concept-Level Reasoning in Scientific Machine Learning

Shusen Liu | 23-ERD-029

Executive Summary

We propose to open the black-box deep learning model by establishing a shared language, i.e., domain concepts, that are understandable for humans while encoded explicitly in the model. By utilizing these concepts as basic building blocks for human reasoning, we can obtain actionable insights from deep learning systems and establish trust in model predictions.

Publications, Presentations, and Patents

Sangwon Jeong, Mingwei Li, Matthew Berger, Shusen Liu, Vidya Setlur, and Chung Haeyong, "Concept Lens: Visually Analyzing the Consistency of Semantic Manipulation in GAN" (Presentation, 2023 IEEE Visualization Conference (VIS), IEEE, Melbourne, Australia, 2023).