Representation learning for robust and adaptable artificial intelligence

The most remarkable trait of deep learning is the ability of deep artificial neural networks to solve any problem they are trained upon according to their own, learnt and bespoke, internal representation of it. This allows such models not to depend anymore on hand-crafted features specified beforehand, and allows modellers to reach for solutions they are not even able to foresee. Such advantage, however, does not come without a price as a lack of control on the inner workings of the model may hinder generalisation or give rise to bizarre idiosyncrasies. Additionally, these models are still far from having attained an intuitive understanding of, and swift adaptability to, new unforeseen problem variations --- a quintessential trait of biological intelligence. In this thesis, we will use the lens of representation learning to explore several failure modes of deep learning models --- sharing a common lack of robustness, flexibility, and adaptation to the unexpected. We will also contextually develop semi-empirical methods to steer model training towards safer, more compliant, predictable, and adaptable behaviour --- by making the role of the internal representation more explicit within the objective being optimised, or by performing specific interventions on it. Applications will include the development of models capable to withstand unforeseen adversarial perturbation of the inputs, to adapt in the multi-task learning of geometrically challenging tasks, and to optimise specific simulated quantum systems within a broad range of configurations.