Wednesday, October 12, 3.15 pm, 2016.

Causal discovery from "big data"
Tom Heskes, Institute for Computing and Information Sciences (iCIS), Radboud University Nijmegen
Abstract: Discovering causal relations from data lies at the heart of most scientific research today. In apparent contradiction with the adagio "correlation does not imply causation", recent theoretical insights indicate that such causal knowledge can also be derived from purely observational data, instead of only from controlled experimentation. In the "big data" era, such observational data is abundant and being able to actually derive causal relationships from very large data sets would open up a wealth of opportunities for improving business, science, government, and healthcare. In this talk, I will sketch how insights from statistics and machine learning may lead to novel approaches for robust discovery of relevant causal relationships.
Location: Visionen
Organizer: Mattias Villani

Wednesday, November 9, 3.15 pm, 2016.

Structured Representation using Latent Variable Models
Cheng Zhang, Disney Research, Pittsburgh and Dept. of Robotics, Perception and Learning at KTH
Abstract: Humans sense the environment, extract information and take actions based on this abstract information. Similarly, machines receive data, extract information and make decisions about unknown variables through inference. Models provide a mechanism for machines to abstract information. This commonly involves learning useful representations which are ideally compact, interpretable and useful for different tasks.
I see representation learning as the task of learning a statistical model of the data where the desired representation is considered as latent parameter. In this talk, I will present my contribution on the design of efficient representation models with focus on structured representation learning with topic models. To learn high quality representations using latent variable models, inference is the key. However, current inference schemes suffer from imbalanced data. I will present a balanced population stochastic variational inference (BP-SVI) scheme for latent representation learning. I have applied these models to different domains including computer vision, robotics, E-health, information retrieval and recommendation systems. In this talk, I will focus on examples within computer vision and E-health.
Location: Visionen
Organizer: Mattias Villani

Wednesday, December 7, 3.15 pm, 2016.

On Machine Learning for Data Privacy
Vicenç Torra, School of Informatics, University of Skövde
Abstract: Data privacy studies and develops methods and tools for avoiding the disclosure of sensitive information. There are three communities working on techical solutions for data privacy. They are the Privacy preserving data mining (PPDM), the privacy enhancing technologies (PETs) and the statistical ddisclosure control (SDC) community.
Definitions of disclosure risk and ways to assess this risk are cornerstones of data privacy. Disclosure risk in data privacy can be seen from two perspectives or dimensions. One dimension focuses on the type of disclosure (identity and attribute discloure) and the other dimension is on the type of measurement (Boolean property or quantitative measurement). Concepts as k-Anonymity, differential privacy, and uniqueness can be seen from this perspective. Our approach to disclosure risk assessment mainly focus on identity disclosure using record linkage (and data integration) to get a quantitative measure of the risk.
In this talk I will give an overview of disclosure risk in data privacy, discuss about record linkage to evaluate the worst case scenario. I will present a supervised learning approach for record linkage. This is a research topic related to metric learning. I will also mention other usages of machine learning in data privacy (e.g., clustering and classification) for the evaluation of information loss and data utility.
Location: Visionen
Organizer: Jose M Pena

The seminars are typically held every fourth Wednesday at 15.15-16.15 in Visionen.
For further information, or if you want to be notified about the seminars by e-mail, please contact Mattias Villani.