Representation learning methods for robotic perception and learning — at the intersection of computational neuroscience and machine learning

Despite the progress made in artificial intelligence and robotics, researchers have yet to fully decode or replicate the mechanisms behind the remarkable ability of the human brain to extract effective and flexible representations from sensor stimuli. Although numerous established machine-learning methods claim inspiration from the brain, uncovering novel concepts from neuroscience can enable further progress in robotics and machine learning. Therefore, this thesis presents research at the intersection of machine learning, robotics, and neuroscience, with an emphasis on representation learning, and perception. First, the thesis introduces work on learning linearly alignable representations using a coupled autoencoder setup. Learning such representations simplifies the computationally demanding task of comparing and aligning probability distributions, a core component of many machine learning methods. Empirical evaluation of the proposed approach demonstrates that it can significantly simplify the solution to the mathematical optimization problem underlying domain adaptation. The core of the thesis focuses on applying principles from neuroscience to improve state-of-the-art representation learning methods. Hand-crafted features and representations learned using multi-modal variational autoencoders and predictive coding are empirically compared in terms of their robustness and data efficiency in navigation and place-recognition tasks in various experiments. Following the superior performance of predictive coding in the performed experiments, this thesis presents a brain-inspired extension to the variational autoencoder framework. Enforcing a slowness prior on latent dynamics in the variational autoencoder facilitates data-efficiency in downstream tasks. Finally, we extend our findings from the domain of representation learning and perception to imitation learning. In the constraint setting of learning from observations only, existing methods are brittle and fail to recover the causal effects of expert actions when access to the target environment is limited. Applying the previously discussed brain-inspired principles to learn representations in state-action spaces solves this problem. The results of the research presented in this thesis indicate that augmenting representation learning methods with principles from neuroscience can help build more data-efficient, robust, and flexible intelligent systems.