The paper discusses the role of Bayesian Deep Learning (BDL) in the current deep learning research landscape. It points out that while most research focuses on predictive accuracy in supervised tasks involving large image and language datasets, there are other overlooked metrics, tasks, and data types that require attention. These include uncertainty, active and continual learning, and scientific data. The authors argue that BDL offers advantages in these diverse settings and can elevate the capabilities of deep learning. They acknowledge existing challenges and highlight potential research avenues to address these obstacles. The discussion concludes by suggesting ways to combine large-scale foundation models with BDL to realize their full potential.
Publication date: 2 Feb 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2402.00809
