Reproduce “End-to-end Feature Alignment: A Simple CNN with Intrinsic Class Attribution” on OpenResearch

About this paper

Parniyan Farvardin, David Chapman

Researchers at the University of Miami developed a Feature-Align CNN (FA-CNN) architecture that generates intrinsically interpretable feature maps by enforcing end-to-end feature alignment. This approach yields penultimate feature maps identical to Grad-CAM saliency maps without additional computation, while achieving classification accuracy comparable to ResNet baselines on CIFAR and ImageNet datasets.

Key insights

• FA-CNN's specific architectural design mathematically guarantees that its penultimate feature maps are identical to Grad-CAM saliency maps, providing these explanations inherently and without additional computation.
• The dampened skip connection ensures that FA-CNN features exhibit a 'Bounded Increment Property through Depth,' enabling the observation of gradual semantic concept evolution layer by layer.
• The end-to-end feature alignment enforced by the architectural components maintains a consistent ordering of semantic concepts from the input pixels to the final class logits.

Summary by alphaXiv.