eXplainable Artificial Intelligence (XAI)

Deep Learning
eXplainable Artificial Intelligence
Concept Extraction
Developing methods for concept extraction, localization, and learning to explain what deep neural networks learn through training.
Published

April 27, 2019

Research Overview

This research focuses on making deep neural networks (DNNs) more transparent and trustworthy through explainable artificial intelligence (XAI), with a particular emphasis on concept extraction. In critical applications such as industrial quality control, understanding what a model has learned and why it makes certain predictions is essential for safety, reliability, and human trust.

We develop methods that generate global explanations for trained neural network models by automatically extracting and localizing human-understandable concepts. Key contributions include:

  • ECLAD (Extracting Concepts with Local Aggregated Descriptors): A method for automatic concept extraction and localization based on pixel-wise aggregations of CNN activation maps, with a rigorous validation framework using synthetic datasets.

  • SPACE (Scale-Preserving Automatic Concept Extraction): An algorithm designed for industrial applications that preserves scale information throughout the concept extraction process, critical for quality control tasks where feature size matters.

  • ECLAD-ts: Extension of concept extraction to the time series domain, enabling the identification and localization of meaningful temporal concepts.

  • Concept Regularization: Closing the loop between concept extraction and model training, using extracted concepts as regularization signals to encourage better-aligned representations.