Interpretable Deep Covolutional Fuzzy Networks

• Author / Creator

  Yeganejou, Mojtaba

• While deep learning has proven to be a powerful new tool for modeling and predicting a wide variety of complex phenomena, those models remain incomprehensible black boxes. This is a critical impediment to the widespread deployment of deep learning technology, as decades of research have found that users simply will not trust (i.e. make decisions based on) a model whose solutions cannot be explained. Fuzzy systems, on the other hand, are by design much more easily understood. We propose to create more comprehensible deep networks by hybridizing them with fuzzy logic. Our proposed architecture first employs a convolutional neural network as an automated feature extractor, and then clusters datasets in that feature space using the Fuzzy C- Means (FCM) and Gustafson and Kessel (GK) clustering algorithms. After hardening the clusters, we employ a fuzzy version of Rocchio's algorithm to classify the data points. We will evaluate our system’s performance on three well-known benchmark datasets (MNIST, Fashion MNIST, and CIFAR-10) by comparing our results with most recent published results (including state-of-the-art). Finally, we will demonstrate the interpretability of the hybrid system on the MNIST dataset.

• Subjects / Keywords

  • XAI
  • Deep Learning
  • Fuzzy Logic

• Graduation date

  Spring 2019

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-42d0-s910

• License

  Permission is hereby granted to the University of Alberta Libraries to reproduce single copies of this thesis and to lend or sell such copies for private, scholarly or scientific research purposes only. Where the thesis is converted to, or otherwise made available in digital form, the University of Alberta will advise potential users of the thesis of these terms. The author reserves all other publication and other rights in association with the copyright in the thesis and, except as herein before provided, neither the thesis nor any substantial portion thereof may be printed or otherwise reproduced in any material form whatsoever without the author's prior written permission.