Neural Networks with Dynamic Recurrent Routing

Neural Network is a powerful Machine Learning technique that demonstrates outstanding performance in such areas as Natural Language Processing, Expert Systems, and Computer Vision. In particular, the recently introduced ResNet model and its variations showcase state-of-the-art results in image classification. ResNet's and most of the previously proposed models' architectures have a fixed structure and apply the same sequence of transformations to all input objects. In this work, we introduce a ResNet-based model ReSet that dynamically selects an individual sequence of transformations for each input object from a trainable set of computational units. Dynamic routing enables the network to learn such a set of meaningful transformations and apply only the required ones to classify each input image. Comparison of our novel model and ResNet-38 showed that the former outperforms the original ResNet on CIFAR-10.1 test set (+1.3% to accuracy). While analyzing the produced paths, we discovered that the proposed ReSet learned to select different routes for images from different classes and similar routes for semantically similar images.