Visualizing the Resilience of Deep Convolutional Network Interpretations

Bhavan Vasu, Andreas E. Savakis

Abstract: This paper aims at visualizing the resiliency of deep net- work interpretations across datasets. We further explore how these interpretations change when network weights are damaged. We utilize Class Activation Maps to obtain heatmaps of deep network interpretations and identify salient local regions. We apply our methods on two remote sensing datasets and demonstrate that representations are resilient across similar datasets. We also demonstrate the benefits of transfer learning for different datasets. We further analyze these interpretations when the network weights are damaged and illustrate that retraining a damaged network is useful in recovering its performance. Our visualization results, based on ResNet50, offer insights in the resiliency of convolutional network architectures.