Topological Interpretation of Deep-Learning Models
Education and Training and Outreach
HPC Community Collaboration
HPC Training and Education
Machine Learning and Artificial Intelligence
TimeSunday, 14 November 202111:15am - 11:45am CST
DescriptionDeveloping trust in the predictions made from an AI-based algorithm is a tantamount concern, especially in systems such as threat detection or medical diagnosis where outcomes may have tragic consequences. This work presents a topologically informed methodology for inferring prominent features in a deep-learning classification model trained on clinical text. Creating a graph of the model's prediction space, we cluster the inputs into the vertices of the graph, and extract subgraphs demonstrating high-predictive accuracy for a given label. These nodes contain a wealth of information about features that the deep-learning model has recognized as important. We demonstrate fidelity of these inferred features to the original model by appending these terms to incorrectly classified documents, and showing an increase in classification accuracy on these augmented documents. This work demonstrates that we may gain actionable insights for subsequent models by improving erroneous classifications and providing a path forward for continuous improvement.