Lecture
Docent lecture Radu-Casian Mihailescu

Despite reaching state-of-the-art performance in relation to key machine learning tasks, deep neural networks are still nowhere close in bridging the gap with respect to the range of essential cognitive abilities associated with human-level intelligence. Although scale has proven to be a definitive driver in building increasingly better performing models, neural nets still exhibit serious vulnerabilities and erroneous behavior in terms of brittleness, spurious correlations, lack of interpretability, or more recently, hallucinating misinformation. In this talk I will set-out by identifying a few key underlying limitations in the deep learning literature, as well as addressing shortcomings of machine learning techniques more broadly.

Following, I will point towards several research directions by reflecting on contributions from my previous and ongoing research. The presentation will focus on approaches towards mitigating some of these identified shortcomings, showcased across various application domains including computer vision, natural language processing and internet of things settings.