Visual Exploration of Multi-Classification Model with High Number of Classes
Authors: Aung Phyo, Khalid Mehtab Khan
Faculty Supervisor: Shahrukh Humayoun
Department: Computer Science
Convolution Neural Network (CNN) approach in Machine Learning (ML) uses combinations of convectional layers and pooling layers to reduce the computational cost and output variations. therefore, it has been applied in different domains from recognizing letters to identify objects in images. However, these all layers typically reside in a black-box, which makes it difficult for ML experts to identify the root cause of problems that may occur during the learning phase such as misclassification of instances. The traditional backtracking solutions are time consuming and hectic for ML practitioners. Targeting this concern, many visual analytics tools have been proposed targeting multi-classification models. However, visual exploration of multi-classification models with large number of classes are not adequately addressed in the literature. Targeting this concern, we present an interactive visual analytics tool that allows users to visual explore multi-classification model, targeting 1K classes in ILSVRC image dataset, using parallel coordinate views in overview+detail style. Users can explore the misclassification (inbound and outbound) cases to get better understanding of problematic classes in the model. Our visual analytics tool provides a Chord diagram view for in-depth inspection of the incorrect classification cases. Several filtering and sorting options are provided at each level to help users in model exploration and inspection.