Visualization-based neural network introspection

Abstract

Artificial intelligence (AI) and the use of neural networks have become omnipresent in recent years. Neural networks model complex mathematical functions that can be based on billions, or even trillions, of parameters. At the same time, neural networks make decisions that can deeply impact our lives. Therefore, understanding, testing, and interpreting these networks and their decisions is an integral part of model development and deployment. While there exist introspection techniques that support such understanding, testing, and interpretation, their adoption for diagnosing systems and explaining decisions can be difficult. Current problems with the adoption of introspection techniques are that they are not easily implemented in one's framework, do not work well in combination to create more meaningful analyses, and are difficult to interpret. Through the integration of existing and novel introspection techniques into visualization interfaces, extensive analysis, actionable insights, and effective diagnosis are made widely available. These visualization interfaces can be incorporated into existing development workflows and are designed to support bespoke analysis needs, which makes the interpretation of findings easier. In this thesis, we present published visualization interfaces in three different areas, namely quality assurance, communication, and AI education. These publications include a visualization approach for correcting mislabeled training data, an interface for automatic network figure generation to communicate network architectures, and an educational environment for recurrent neural networks (RNNs). Finally, to unify the diverse landscape of AI visualization interfaces, we also present a framework for composing, reusing, exploring, and sharing such interactive machine learning (ML) interfaces.