Yue Ren, Swansea University
Ximena Fernandez, Swansea University
15 December 2020
Department of Mathematics, University of York, on Remo and Zoom (details will be sent to the AAG mailing list)
Emilie Dufresne, University of York
13:30-14:00 Welcome (on Remo)
14:00-14:45 Ximena Fernandez: Geometric and Topological Inference for Data Analysis (on Zoom)
14:45-15:15 Break (on Remo)
15:15-16:00 Yue Ren: Tropical varieties of neural networks (on Zoom)
Abstract: In this talk we approach the problem of learning information about a geometric object from a finite set of (possibly noisy) sample points drawn respect to some unknown distribution. More concretely, given a smooth manifold and a density that produces the sample, we consider an intrinsic density-based metric, known as the Fermat distance. We construct a computable distance over the sample and prove that this sample metric space is a good estimator of the manifold (in the sense of Gromov-Haussdorf). Finally, we present some applications of this result in topological data analysis, showing how this approach outperforms more standard methods with computational experiments in synthetic and real datasets.
Abstract: In this talk, we introduce tropical varieties arising from neural networks with piecewise linear activations, and discuss how their geometry affects their expressivity. In particular, we will use Weibel's f-Vector Theorem to derive optimal bounds for single-layered maxout networks, and Speyer's f-Vector Theorem to analyse networks with heavily restricted weights. We conclude with an initializing strategy for maxout networks based on our results.
------------------------------------------------------------------------
We are grateful for the financial support from the Glasgow Mathematical Journal Learning and Research Support Fund, from the Edinburgh Mathematical Society, the London Mathematical Society.