Emanuel Ström: Error Bounds for Deep-Learning

Deep learning is one of those elusive terms that floats around in conversations about modern AI advancements (chat GPT, stable diffusion, Alpha GO, DLDSR etc. all use it). A common argument against using Deep-Learning methods is that there is very little understanding of how they work. While true, the claim is sometimes repeated almost like a mantra -- that conveniently omits any information about our current best understanding of deep learning models. My goal with this talk is to give insight into one specific research direction in the field. The talk will be structured as follows:

(1) I will formulate deep learning applied to regression, 

(2) discuss error contributions and common ways of bounding them, and

(3) Present and prove (at least partly) an important theorem that is less than 5 years old.