An Interactive 3D Visualization Tool to Analyze the Functionality of Convolutional Neural Networks

• Author / Creator

  Fatemi Pour, Farnoosh

• Convolutional Neural Networks have outperformed traditional image processing approaches for many image classification tasks. However, despite achieving high classification accuracy on multiple datasets, these machine learning approaches are not transparent and act more like a black box. For this reason, explaining how they work and how they perform classifications is essential for research and education.

  This lack of interpretability prevents students and beginners in the field from fully understanding the internal functionality of these networks. Moreover, it makes it hard for an expert to debug the models and enhance their performance. Lastly, not explaining how a model produces an output decreases the user's trust in such machine learning models, especially in medicine.

  Visualizing the internal variables of convolutional networks enables us to understand how the model processes the input data and generates the output classifications. It also allows us to show how neural networks work by exploring the sensitivity of the model to small changes in the input and how the internal variables affect the results of the classification.

  These models usually have multiple convolutional and fully-connected layers with hundreds or even thousands of neurons in each layer. Visualizing this large number of variables intuitively is challenging as it is computationally expensive and hard to display with limited screen resolution.

  This thesis proposes a novel interactive tool to visualize convolutional neural networks in using 3D graphics. The first part of our method partitions the feature maps into clusters instead of showing all feature maps simultaneously. We use a K-Means clustering algorithm to cluster the feature maps and then use principal component analysis to represent each cluster's parameters. The tool also allows the user to create modified versions of the original input by applying binary masks and helps them to analyze the impact of removing different regions of the initial input data.

  We evaluate this new tool by surveying eight participants with diverse backgrounds and asking them multiple questions about its effectiveness and user experience. In addition, we ask participants to compare the proposed tool with another baseline visualization method. Our evaluation indicates that clustering the feature maps in each layer is an effective and helpful way for the user to understand how convolutional neural networks work.

• Subjects / Keywords

  • Visualization
  • Convolutional Neural Networks
  • Visual Analytics

• Graduation date

  Fall 2022

• Type of Item

  Thesis

• Degree

  Master of Science

• DOI

  https://doi.org/10.7939/r3-ezxq-1e42

• License

  This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.