Leveraging Latent Fields for Accurately Attributing Model Behavior




Payne, Ethan Robert

Journal Title

Journal ISSN

Volume Title



Large Machine Learning and Artificial Intelligence models have undeniably become essential tools for addressing many modern problems, so the need for explainability and transparency with respect to such models is now greater than ever. Many existing methods for computing attributions lack a strong underlying explanation for their utility, and are inflexible with respect to differing user requirements. Furthermore, many existing approaches to problems in computer vision fail to efficiently utilize the information present in the data, or fail to take advantage of opportunities for a problem-centric design of solutions. This work proposes several approaches increasing the explainability of existing models, and for approaching new problems from a perspective of transparent design. The primary contribution of this work is a novel formulation of integrated attributions for generating informative statistics with respect to model inputs and model parameters. These Generalized Integrated Attributions provide a transparent means of extracting diverse sources of information regarding high-dimensional input and parameter spaces, resulting in improved interpretability as well as increased utility for applications such as strategic training and unlearning. Additionally, this work describes methods for increasing data efficiency in model training schemes, and identifies several opportunities for explainable design in addressing common computer vision problems. All causal explanations are inherently subjective, and no tool will ever be guaranteed to perform perfectly as intended, but by internalizing the principles of explainability, transparency, and interpretability, we can more quickly and efficiently develop statistics and models which are more useful and reliable.



Attribution, Causal Inference, Computer Vision, Explainability, Interpretability, Transparency



Computer Science