Ivan Dungan Research Focus
Research Focus: Dr. Ivan Dungan
Associate Professor, Department of Mathematics
Francis Marion University
Dr. Ivan Dungan is an Associate Professor in the Department of Mathematics at Francis Marion University (FMU) and is a member of Thrust 1: XAI-Enabled Biomedical Devices of the ADAPT in SC Project. His research interests are rooted in Topological Data Analysis (TDA), Probabilistic Graphical Models (PGM), and Causal Discovery with the general motivation of developing Trustworthy and Explainable Artificial Intelligence (XAI). Besides his own research interests in PGMs, their intuitive interpretation and ability to be used for a variety of applications and data types, make them especially accessible to undergraduate research students and a gateway to understanding more complex methods in Artificial Intelligence (AI). With the support of the ADAPT in SC Award, FMU, and companies, Dr. Dungan has been able to involve many undergraduate students at FMU in his research of XAI methods.
Causal Discovery algorithms in AI attempt to learn causal relationships between features in multimodal data. A user such as a clinician or patient with access to an AI model that combines causal relations and predictions could navigate the relations to reason and explain a prediction much like humans. This observation supports developing Causal Discovery algorithms and applying them to different multimodal data sets for XAI.
Dr. Dungan and his collaborators in Thrust 1 have developed an XAI method based on causal discovery which can be applied to image classification. A traditional PGM method for image classification used the Scale-Invariant Feature Transform (SIFT) and the Naïve Bayes Nearest Neighbor (NBNN) classifier. The NBNN approach uses a PGM to characterize an image relative to a ‘memory’ of training images. The NBNN classifier is limited in expressiveness due to network assumptions and is ill-suited for causal discovery since the method characterizes descriptors of a particular image, not general features. Our approach clusters the image descriptors of the training images into features and learns the appropriate PGM, ideally a Causal Network, on these learned features using the well-established Causal Discovery method called the Peter-Clark (PC) Algorithm. This approach increases the expressiveness and the interpretability of its results.
A visual of each image can be generated to ‘explain’ the classification by identifying the clustered features of the image, independent features, and importance of the necessary features. We have done preliminary testing on a wound image data set and are currently comparing its image classification performance with Convolutional Neural Networks (CNN). We also are leveraging the method’s ability to be implemented on multimodal data to include clinical data and data from the biosensor created by Dr. Jordon Gilmore of Clemson University and Thrust 1 co-leader.
Dr. Dungan is also researching on the theoretical development of Causal Discovery. In particular, he and his collaborators are researching advancements in the PC algorithm. These advancements are for both static and dynamic PGMs.