1:30–5:30 pm
Stevanovich Center, MS 112
5727 South University Avenue
MONDAY, JUNE 17, WEDNESDAY, JUNE 19, and FRIDAY, JUNE 21, 2019
Stevanovich Center, MS 112
5727 S. University Avenue
Chicago, Illinois 60637
Monday, June 17, 2019, at 1:30-2:30 PM, and 3:00-4:00 PM
Lectures 1 & 2: Latent factor models
Speaker: David Bindel, Cornell University
Approximate low-rank factorizations pervade matrix data analysis, often interpreted in terms of latent factor models. After discussing the ubiquitous singular value decomposition (aka PCA), we turn to factorizations such as the interpolative decomposition and the CUR factorization that offer advantages in terms of interpretability and ease of computation. We then discuss constrained approximate factorizations, particularly non-negative matrix factorizations and topic models, which are often particularly useful for decomposing data into sparse parts. Unfortunately, these decompositions may be very expensive to compute, at least in principal. But in many practical applications one can make a separability assumption that allows for relatively inexpensive algorithms. In particular, we show how to the separability assumption enables efficient linear-algebra-based algorithms for topic modeling, and how linear algebraic preprocessing can be used to “clean up” the data and improve the quality of the resulting topics.
Monday, June 17, 2019, at 4:30-5:30 PM
Seminar I: Empirical risk minimization over deep neural networks overcomes the curse of dimensionality in the numerical approximation of Kolmogorov PDEs
Speaker: Julius Berner, University of Vienna
Recently, methods based on empirical risk minimization (ERM) over deep neural network hypothesis classes have been applied to the numerical solution of PDEs with great success. We consider under which conditions ERM over a neural network hypothesis class approximates, with high probability, the solution of a d-dimensional Kolmogorov PDE with affine drift and diffusion coefficients up to error e. We establish that such an approximation can be achieved with both the size of the hypothesis class and the number of training samples scaling only polynomially in d and 1/e.
