Convolutional Dynamic Alignment Networks for Interpretable Classifications

TL;DR

Convolutional Dynamic Alignment Networks (CoDA-Nets) are interpretable neural classifiers that use dynamic input-dependent linear transformations to produce high-quality, visually meaningful contribution maps while maintaining competitive accuracy.

Contribution

This paper introduces CoDA-Nets, a novel neural network architecture with inherent interpretability through dynamic alignment units that enable linear decomposition of predictions.

Findings

CoDA-Nets outperform existing attribution methods in quantitative metrics.

They achieve classification accuracy comparable to ResNet and VGG models.

Contribution maps from CoDA-Nets align well with discriminative input patterns.

Abstract

We introduce a new family of neural network models called Convolutional Dynamic Alignment Networks (CoDA-Nets), which are performant classifiers with a high degree of inherent interpretability. Their core building blocks are Dynamic Alignment Units (DAUs), which linearly transform their input with weight vectors that dynamically align with task-relevant patterns. As a result, CoDA-Nets model the classification prediction through a series of input-dependent linear transformations, allowing for linear decomposition of the output into individual input contributions. Given the alignment of the DAUs, the resulting contribution maps align with discriminative input patterns. These model-inherent decompositions are of high visual quality and outperform existing attribution methods under quantitative metrics. Further, CoDA-Nets constitute performant classifiers, achieving on par results to ResNet and VGG models on e.g. CIFAR-10 and TinyImagenet.