Please join the webinar here.

Al and Bob are very similar, with one specific exception. This exception might pertain to their age, annual salary, health, height, education or sports team preference. Whatever it is, this difference is non…

Online optimization is a powerful framework in machine learning that has seen numerous applications to problems in energy and sustainability. In my group at Caltech, we began by applying online optimization to ‘right-size’ capacity in data centers nearly a decade ago; and by now tools from online optimization have been applied to develop algorithms for geographical load balancing among data centers, demand response, generation planning, energy storage management, and beyond. In this talk, I will highlight both the applications of online optimization and the theoretical progress that has been driven by applications in energy and sustainability. Over a decade, we have moved from designing algorithms for one-dimensional problems with restrictive assumptions on costs to general results for high-dimensional, non-convex problems that highlight the role of constraints, predictions, delay, multi-timescale control, and more. 

Nowadays, all sorts of sensors, from ground to space, collect a huge volume of data about the Earth. Recent advances in artificial intelligence (AI) provide unprecedented opportunities for data-driven hydrological modeling using such “Big Earth Data”. However, many critical issues remain to be addressed. For example, there lacks efficient frameworks for data-driven modeling under the big data condition. Moreover, with the amazing predicting power of deep learning (DL) demonstrated in the field of hydrology, whether and how hydrology theory can still play an important role in hydrological modeling is under debate. This presentation introduces two recent studies conducted by Prof. Yi Zheng’s group which attempt to address the above issues. A novel approach to data-driven hydrological modeling was first developed.

We present new theoretical results for procedures identifying coarse structures in a given data set by means of appropriate spectral embeddings. We combine ideas from spectral geometry, metastability, optimal transport, and spectral analysis of weighted graph Laplacians to describe the embedding geometry. Our analysis focuses on 1) studying the…

Multiple organizations often wish to aggregate their sensitive data and learn from it, but they cannot do so because they cannot share their data. For example, banks wish to run joint anti-money laundering algorithms over their aggregate transaction data because criminals hide their traces across different banks.

Bio:
Raluca Ada Popa is an assistant professor of computer science at University of California, Berkeley working in computer security, systems, and applied cryptography. She is a co-founder and co-director of the RISELab at UC Berkeley, as well as a co-founder and CTO of a cybersecurity startup called PreVeil. Raluca received her doctoral degree in computer science as well as her master’s and two bachelors’ degrees from Massachusetts Institute of Technology. She is the recipient of a Sloan Foundation Fellowship award, NSF Career, Technology Review 35 Innovators under 35, Microsoft Faculty Fellowship, and a George M. Sprowls Award for best MIT computer science doctoral thesis.

To request accommodations for a disability please contact Jean Butcher, [email protected], at least one week prior to the event.

In this talk, we propose an analysis of gradient descent on wide two-layer ReLU neural networks that leads to sharp characterizations of the learned predictor. The main idea is to study the dynamics when the width of the hidden layer goes to infinity, which is a Wasserstein gradient flow. While this dynamics evolves on a non-convex landscape,…

Abstract:            The computation of extremal eigenvalues of large, sparse matrices has proven to be one of the most important problems in numerical linear algebra. Krylov subspace methods are a powerful class of techniques for this problem, most notably the Arnoldi process for non-symmetric matrices and the…

Deep network approximation is a powerful tool of function approximation via composition. We will present a few new thoughts on deep network approximation from the point of view of scientific computing in practice: given an arbitrary width and depth of neural networks, what is the optimal approximation rate of various function…

This webinar will provide several practical considerations to help you better manage your research data between the points of collection and analysis. We will review the principles of open research and cover best practices for documentation and metadata generation amidst collation, aggregation, and cleaning tasks. Then we will survey some…

Standard machine learning produces models that are highly accurate on average but that degrade dramatically when the test distribtion deviates from the training distribution. While one can train robust models, this often comes at the expense of standard accuracy (on the training distribution). We study this tradeoff in two…