Berlin 2018 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

MM: Fachverband Metall- und Materialphysik

MM 66: Topical Session (Symposium MM): Big Data in Materials Science - Managing and exploiting the raw material of the 21st century

MM 66.1: Topical Talk

Donnerstag, 15. März 2018, 17:30–18:00, H 0107

Discovering Interpretable Patterns, Correlations, and Causality — •Jilles Vreeken — Max Planck Institute for Informatics, Saarbrücken, Germany — Saarland University, Saarbrücken, Germany

To gain non-trivial insight from data using machine learning, we need to be able to interpret what these results mean. This we can either do by staring long and hard at the highly complex and non-linear models that methods such as support vector machines or deep learning provide when we run them on our data. This most often ends in us throwing the towel, as these models are extremely difficult to understand. Alternatively, we can require the learning method to report in a language we can (much) (more) easily understand, instructing it to discover things beyond what we already know.

In this talk, I will give an introduction to this latter, interpretable approach. In particular, I will explain the power of local modeling, that of non-parametric correlation discovery, that of pattern languages, will give examples of recent discoveries we made on materials science data using a technique called subgroup discovery, and an outlook on very recent approach to discover causal dependencies in data without having to make (almost) any assumptions.