Dr. Emily Stark (University of Utah)
Filtered ends and obstructing group actions abstract
Abstract:
Studying the topology of a space at infinity offers a powerful perspective in geometric group theory. Filtered ends capture a space at infinity relative to a subspace: one considers complements of increasing neighborhoods of a subcomplex in a simplicial complex. The i-th homology groups of the complements form an inverse system, and the i-th Cech homology group is the associated inverse limit. The goal is then to compute the Cech homology groups of the filtered end and to use homological arguments to study these pairs of spaces. In this talk, I will explain how both goals are possible in the setting of coarse embeddings into coarse PD(n) spaces. Indeed, Kapovich--Kleiner proved a coarse Alexander Duality theorem to compute these deep homology groups. We extend their construction to prove a relative duality theorem, thus developing the homology theory further. As applications, one can use the deep homology groups together with homological arguments to prove that certain groups cannot act properly on a given manifold. This is joint work with Chris Hruska and Hung Cong Tran.
Prof. Dr. Andrea Montanari (Stanford University, Stanfort, USA)
The generalization error of overparametrized models: Insights from exact asymptotics abstract
Abstract:
In a canonical supervised learning setting, we are given n data samples, each comprising a feature vector anda label, or response variable. We are asked to learn a function f that can predict the the label associatedto a new --unseen-- feature vector. How is it possible that the model learnt from observed data generalizes tonew points? Classical learning theory assumes that data points are drawn i.i.d. from a common distribution and arguethat this phenomenon is a consequence of uniform convergence: the training error is close to its expectation uniformlyover all models in a certain class.Modern deep learning systems appear to defy this viewpoint: they achieve training error that is significantly smaller than the test error,and yet generalize well to new data. I will present a sequence of high-dimensional examples in which this phenomenon can beunderstood in detail.[Based on joint work with Song Mei, Feng Ruan, Youngtak Sohn, Jun Yan]