What is it about?
Using entropy and Jensen-Shannon Divergence to measure class-specificity at each layer of a Convolutional Neural Network. Deeper layers in a CNN learn more complex and class-specific features, which helps in distinguishing between classes in image classification
Featured Image
Photo by Alina Grubnyak on Unsplash
Why is it important?
While previous methods exist for explaining image classifiers, this research offers a new perspective by analyzing how feature value distributions differ between classes across network layers. This approach provides a novel way to understand how CNNs learn and transition from general to class-specific representations.
Perspectives
1. We feed a batch of 400 samples from each of two classes to a trained VGG-19 two-class classifier to estimate the two class probabilities at each location in each layer tensor. 2. The JSD quantifies the dissimilarity between the two probabilities. 3. A layer is class agnostic if the differences between the two classes are small and uniformly so over the whole tensor. A class specific layer would have some, not necessarily all, features with larger differences. 4. The Interquartile Range (IQR), which is more resistant to extreme values than, say, the standard deviation, is used to measure the variation of the dissimilarities in each layer.
Debanjali Banerjee
Earlham College
Read the Original
This page is a summary of: Information Flow in Deep Learning Classification Networks, March 2025, ACM (Association for Computing Machinery),
DOI: 10.1145/3672608.3707970.
You can read the full text:
Contributors
The following have contributed to this page
