Convolutional Dynamic Alignment Networks for Interpretable Classifications

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. The core building blocks are Dynamic Alignment Units (DAUs) which are "dynamic linear" (i.e., input-dependent linear) and align their weight vectors with task-relevant input patterns during optimisation. As a result, CoDA-Nets model the classification prediction through a series of dynamic linear transformations, which allows for linear decomposition of the prediction into individual input contributions. Due to the alignment property of the DAUs, the resulting contribution maps align with discriminative input patterns. These model-inherent contribution maps are of high visual quality and outperform existing attribution methods under quantitative metrics. Further, our architectures constitute performant classifiers, achieving on par results to models from the ResNet and VGG model families e.g. for CIFAR-10 and TinyImagenet.