Improving the Accuracy and Robustness of CNNs Using a Deep CCA Neural Data Regularizer

  • Author / Creator
    Pirlot, Cassidy
  • Convolution Neural Networks (CNNs) have rapidly evolved since their neuroscience beginnings. These models efficiently and accurately classify images by optimizing the model’s hidden representations to these images through training. These representa- tions have been shown to resemble neural data from the primate visual system as the accuracy of the model improves. Works have been produced to exploit these findings to examine if the more (mammalian) brain-like a model’s hidden representations are, the more (mammalian) brain-like the model’s performances will be. Further, perfor- mance from a model that is human-like would achieve high accuracy, high super-class accuracy, and robustness. We expand on this work by using a neural data (ND) regularizer that uses Deep Canonical Correlation Analysis (DCCA). The regularizer optimizes the resemblance between the CNN’s hidden representations to an image and the representations found in the mammalian visual pathway to the same image. Compared to CNNs without the ND regularizer, the ND regularized CNN resulted in higher accuracy and super-class accuracy, as well as becoming more robust to ad- versarial examples. These outcomes provide evidence that pushing CNNs to become more brain-like is not only achievable, but will also result in a better performing model.

  • Subjects / Keywords
  • Graduation date
    Fall 2022
  • Type of Item
  • Degree
    Master of Science
  • DOI
  • 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.