The Bayesian approach to deep learning aims to encode prior knowledge into models, but defining priors is often challenging. These challenges are discussed in "Leveraging Connections Between Deep Architectures and Bayesian Nonparametrics", a talk by Martin Trapp of Aalto University. The talk is part of OFAI's 2023 Winter/Spring Lecture Series.
Members of the public are cordially invited to attend the talk via Zoom on Wednesday, 18 January at 18:30 CET (UTC+1):
URL: https://us06web.zoom.us/j/84282442460?pwd=NHVhQnJXOVdZTWtNcWNRQllaQWFnQT09
Meeting ID: 842 8244 2460
Passcode: 678868
Talk abstract: Deep architectures have become an integral part of modern AI systems. However, despite their computational benefits, concerns about their robustness and missing interpretability often limit their applicability in scenarios where trustworthiness is of importance. This motivates the Bayesian approach to deep learning, which aims to encode prior knowledge into the model. However, defining Bayesian priors for deep architecture is often challenging. In this talk, I will discuss how drawing connections to Bayesian nonparametric priors can help in (i) encoding conservative behaviour into deep learning models, (ii) understanding priors over deep tractable models, and (iii) developing tools for uncertainty quantification in computer vision tasks.
Speaker biography: Martin Trapp is an Academy of Finland postdoctoral researcher at Aalto University, Finland. Previously, Martin was a postdoctoral researcher working with Arno Solin at Aalto University and finished his PhD at the Graz University of Technology under Franz Pernkopf and Robert Peharz. During his PhD, Martin worked as a researcher at the Austrian Research Institute for Artificial Intelligence and was a visiting researcher at the Computational and Biological Learning Lab at the University of Cambridge. His main research interests are in the intersection of deep learning, probabilistic circuits, probabilistic programming, and Bayesian nonparametrics, with the goal of building systems that are computationally efficient, reliable, and can adjust their complexity to the data distribution.
