Public defence in Automation, Systems and Control Engineering, M.Sc. Oliver Struckmeier

The title of the thesis: Representation learning methods for robotic perception and learning - at the intersection of computational neuroscience and machine learning 

Doctoral student: Oliver Struckmeier
Opponent: Prof. Jaakko Peltonen, Tampere University, Finland
Custos: Prof. Ville Kyrki, Aalto University School of Electrical Engineering, Department of Electrical Engineering and Automation

In recent years, advances in artificial intelligence have fundamentally transformed the way we interact with technology. Despite these advancements, intelligent systems still struggle with acquiring a holistic understanding of the physical world — a crucial ability for learning and adapting, which is essential for meaningful interaction with this environment. Recent advances in the field of neuroscience have advanced our understanding of how sensory processing processes occur in the brain, which promises improvements when building intelligent systems. 

First, this dissertation discusses methods for measuring and comparing the quality of learned representations. Following this discussion, the positive impact of learning linearly alignable representations on performance and data efficiency in domain adaptation tasks is investigated. Next, three case studies are presented, in which principles from neuroscience have been applied to improve the robustness and data efficiency of state-of-the-art machine learning methods. Most notably leveraging temporal similarity and prediction of sensory information is used to solve problems like efficient and robust downstream task-learning and imitation learning. In a set of experiments, the explored concepts are implemented on neuromorphic hardware and validated in navigation tasks. 

The findings presented in this dissertation affirm the potential of brain-inspired representation learning methods. Specifically, prediction and temporal slowness are concepts that have shown significant impact when applied to state-of-the-art representation learning methods. In conclusion, brain-inspired representation learning presents a promising direction toward designing more efficient, flexible, and robust intelligent systems.

Keywords: Representation Learning, Imitation Learning, Neurorobotics

Thesis available for public display 10 days prior to the defence at: https://aaltodoc.aalto.fi/doc_public/eonly/riiputus/

Contact:

Doctoral theses in the School of Electrical Engineering: https://aaltodoc.aalto.fi/handle/123456789/53