University of Twente Student Theses


Extending explainability of generative classifiers with prototypical parts

Peters, M.D.J. (2022) Extending explainability of generative classifiers with prototypical parts.

[img] PDF
Abstract:The field of Explainable Artificial Intelligence is a response to the increase in use of Artificial Intelligence and lack of insights into the reasonings behind a model’s decisions. Prototypical parts have been suggested as explanations that are easy to interpret by humans, and generative models as being able to give more insight into the classification of Out-of-Distribution data. We introduce a model, ProtINN, that combines a generative model with prototypical part explanations, to give insightful explanations on its predictions. ProtINN segments the input images and uses a pre-trained Invertible Neural Network to get the segments’ activations. These activations are used for clustering to create prototypes, and to calculate the similarity between the prototypes and the segments of the to-be classified images. These similarity scores are then translated to class prediction scores by a single linear layer. To explain the class prediction scores, we present visualisations which show the prototypes, similarity scores, prediction layer weights, and input image segments. Our experiments on the ImageNet dataset show that, while we have some loss in accuracy, we gain a lot of insights into the predictions with our visualisations.
Item Type:Essay (Master)
Faculty:EEMCS: Electrical Engineering, Mathematics and Computer Science
Subject:54 computer science
Programme:Computer Science MSc (60300)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page