Universal approximation with neural networks through the lens of computability
| Quand ? |
Le 21/03/2023, de 13:00 à 14:00 |
|---|---|
| Participants |
Laura Thesing |
| Ajouter un événement au calendrier |
vCal
iCal
|
Speaker: Laura Thesing (Postdoctoral researcher - LMU Munich)
Title: Universal approximation with neural networks through the lens of computability
Abstract:
The impact deep learning and artificial intelligence have on our daily lives, and society can hardly be overestimated. The use cases vary from health care to migration and law decisions. Especially in these highly sensitive areas, the robustness of the methods is of utmost importance. The problem of robustness has been highlighted in a variety of publications on adversarial examples and hallucinations of AI systems.
There are recent results showing the existence of tasks, for example in inverse problems, which are non-computable with neural networks. Which means that no accurate solution can be obtained independent on the learning mechanism. Therefore, we want to discuss in this talk for which function classes a correct approximation with neural networks is possible on a computing machine. To do so we relate the family of computable functions to those which are computable with a neural network approximation. Moreover, in the spirit of generalized hardness of approximation we do not aim for infinite accuracy but ask for eps_0 such that the solution for eps > eps_0 can be found and that a solution for higher accuracies becomes intractable.
I introduce all information about computability which are needed such that no background knowledge is required.
More information: https://scholar.google.com/citations?hl=fr&user=GJmfKawAAAAJ&view_op=list_works&sortby=pubdate
Talk in room M7 101 (Campus Monod, ENS de Lyon)