Introduces a CNN architecture where feature maps are mathematically identical to Grad-CAM saliency maps by design, rather than post-hoc.
March 30, 2026
Original Paper
End-to-end Feature Alignment: A Simple CNN with Intrinsic Class Attribution
arXiv · 2603.25798
The Takeaway
It moves interpretability from an 'afterthought' (post-hoc attribution) to an intrinsic property of the architecture. Practitioners can now build models that are transparent by design without sacrificing performance on standard benchmarks.
From the abstract
We present Feature-Align CNN (FA-CNN), a prototype CNN architecture with intrinsic class attribution through end-to-end feature alignment. Our intuition is that the use of unordered operations such as Linear and Conv2D layers cause unnecessary shuffling and mixing of semantic concepts, thereby making raw feature maps difficult to understand. We introduce two new order preserving layers, the dampened skip connection, and the global average pooling classifier head. These layers force the model to