Convolutional Dynamic Alignment Networks for Interpretable Classifications

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