Augmenting Classic Algorithms with Neural Components for Strong Generalisation on Ambiguous and High-Dimensional Data

We augment classic algorithms with learned components to adapt them to domains currently dominated by deep learning models. Two traditional sorting algorithms with learnable neural building blocks are applied to visual data with apriori unknown symbols and rules. The models are quickly and reliably trained end-to-end in a supervised setting. Our models learn symbol representations and generalise better than generic neural network models to longer input sequences.