End-to-end Feature Alignment: A Simple CNN with Intrinsic Class Attribution

Researchers Parniyan Farvardin and David Chapman have published a paper introducing the Feature-Align CNN (FA-CNN), a novel convolutional neural network architecture that builds interpretability directly into its design. The core innovation addresses a fundamental problem in modern vision models: standard operations like Linear and Conv2D layers create "unnecessary shuffling and mixing of semantic concepts," making intermediate feature maps difficult to understand. FA-CNN solves this by introducing two new order-preserving layers—a dampened skip connection and a global average pooling classifier head—that enforce a strict, end-to-end alignment of features from the raw input pixels all the way to the final class predictions.

This architectural constraint has a profound theoretical result: the authors prove that FA-CNN's penultimate feature maps are mathematically identical to the saliency maps produced by Grad-CAM, a popular post-hoc explanation technique. This means the model's internal reasoning is transparent by construction, eliminating the need for separate, often unreliable, explanation tools. Furthermore, the paper demonstrates that these aligned feature maps evolve gradually layer-by-layer, visually tracing how the network transforms input features into class-specific activations. While maintaining competitive accuracy on benchmark datasets, FA-CNN's intrinsic attribution was validated in a "percent pixels removed" interpretability task, where its raw feature maps were compared favorably against Grad-CAM and permutation-based methods.