• deep learning
• machine learning
• machine reasoning
• representation learning

Finding effective ways to reach an integration between learning and reasoning within deep neural networks is a long-standing problem of AI. A major breakthrough in this area has the potential of bringing the capabilities of current machine learning systems to the next level, making deep neural networks able to tackle a whole new range of problems, in a way that would have been unthinkable just a few years ago. This thesis explores possible ways to enrich the internal representations of deep learning models, under the long-term perspective of finding good inductive biases for supporting a smooth integration between learning and reasoning. The contributions presented in this document approach the problem from three different directions. First, a novel way to structure the latent representation of deep neural networks is introduced, allowing such representations to disentangle the different generative factors underlying the data. Then, a technique to enrich the same latent representations with external prior information is described, demonstrating its application on the challenging task of automatic music generation. Finally, a new benchmark for accurately measuring the systematic generalization capabilities of reasoning models is presented, based on the prediction of stoichiometrically-balanced chemical reactions. We hope that this thesis could give an in-depth overview of the current research in the fields of representation learning and learning/reasoning integration, as well as making some noteworthy contributions to the research community.