ETH-FDS seminar series

More information about ETH Foundations of Data Science can be found here

Modal title

Modal content

Please subscribe here if you would you like to be notified about these presentations via e-mail. Moreover you can subscribe to the iCal/ics Calender.

Spring Semester 2020

Date / Time

Speaker

Title

Location

22 January 2020
13:15-14:15

Uri Shalit
Technion, Haifa

Event Details

ETH-FDS seminar

Title	Causality-inspired machine learning
Speaker, Affiliation	Uri Shalit, Technion, Haifa
Date, Time	22 January 2020, 13:15-14:15
Location	HG E 1.1
Abstract	We will present three recent projects where ideas from causal inference have inspired us to find new approaches to problems in machine learning. First, we will see how the idea of negative controls led us to find a way to perform off-policy evaluation in a partially observable Markov decision process (POMDP). We will then present how using the idea of independence of cause and mechanism (ICM) can be used to help learn predictive models that are stable against a-priori unknown distributional shifts. Finally, we will see how thinking in terms of causal graphs led us to a new method for learning computer vision models that can better generalize to unseen object-attribute compositions in images.
Assets	Video Uri Shalit - ETH-FDS talk on 22 January 2020file_download

Causality-inspired machine learningread_more

HG E 1.1

2 March 2020
12:15-13:15

Mário A. T. Figueiredo
Instituto Superior Técnico, Universidade de Lisboa, Portugal

Event Details

ETH-FDS seminar

Title	Selection and Clustering of Correlated Variables
Speaker, Affiliation	Mário A. T. Figueiredo, Instituto Superior Técnico, Universidade de Lisboa, Portugal
Date, Time	2 March 2020, 12:15-13:15
Location	HG D 7.1
Abstract	Linear (and generalized linear) regression (LR) is an old, but a still essential, tool in statistical data science: its goal is to learn to predict a (response) variable as a linear combination of other (explanatory) variables. A central problem in LR is the selection of relevant variables, an important task for several reasons: using fewer variables tends to yield better generalization; identifying the relevant variables may be meaningful (e.g.,identifying which gene expressions are relevant to predict a certain disease). In the past quarter-century, variable selection based on sparsity- inducing regularizers has been a central paradigm, the most famous of these regularizers being the LASSO, which has been intensively studied, extended, and applied. In high-dimensional problems, it is natural to have highly-correlated variables. For example, it is common that several genes are strongly correlated (co-regulated), thus simultaneously relevant as predictors of some response. In this case, sparsity-inducing regularization may fail: it may select an arbitrary subset of correlated variables; it is unstable (the selected variables may change drastically, with only small changes in the data). However, in many applications, it is desirable to identify all the relevant variables, rather than an arbitrary subset thereof, a desideratum for which several approaches have been proposed. This talk will be devoted to a recent class of regularizers for this goal, called ordered weighted l1 (OWL). The key feature of OWL is that it is able to explicitly identify (i.e. cluster) sufficiently-correlated features, without the need to actually compute these correlations. Several theoretical results characterizing OWL will be presented, including connections to the mathematics of economic inequality and with “robustified” regression. Computational and optimization aspects will also be addressed, as well as recent applications in subspace clustering, learning Gaussian graphical models, and deep neural networks.
Assets	Video Mário Figueiredo - ETH-FDS talk on 2 March 2020file_download

Selection and Clustering of Correlated Variablesread_more

HG D 7.1

18 May 2020
10:00-11:00

Ming Yuan
Institute for Theoretical Studies, ETH Zurich and Columbia University

Event Details

ETH-FDS seminar

Title	e-Seminar: Information Based Complexity of High Dimensional Sparse Functions
Speaker, Affiliation	Ming Yuan, Institute for Theoretical Studies, ETH Zurich and Columbia University
Date, Time	18 May 2020, 10:00-11:00
Location	Zoom Lecture
Abstract	We investigate the optimal sample complexity of recovering a general high dimensional smooth and sparse function based on point queries. Our result provides a precise characterization of the potential loss, or lack thereof, in information when restricting to point queries as opposed to the more general linear queries, as well as the benefit of randomization and adaption. In addition, we also developed a general framework for function approximation to mitigate the curse of dimensionality that can also be easily adapted to incorporate further structure such as lower order interactions, leading to sample complexities better than those obtained earlier in the literature.
Assets	Video Ming Yuan - ETH-FDS talk on 18 May 2020file_download

e-Seminar: Information Based Complexity of High Dimensional Sparse Functionsread_more

Zoom Lecture

Archive: SS 24 AS 23 SS 23 AS 22 SS 22 AS 21 SS 21 AS 20 SS 20 AS 19 SS 19