"Visual Analysis for Early Data Discoveries"

Hendrik Strobelt, IBM

Thursday, November 16, 2017 at 12:00 Noon

Room 227 (CIT - 2nd Floor)

Data Visualization is effectively used to explore and explain complex data. In the talk, I will present three visual tools for different data domains that can help to formulate hypotheses in early stages of data discovery:

Data often comes with categorical information that assigns each item to one or multiple sets. We developed UpSet to explore the set intersections. Interactively, a user can filter and recombine intersections and calculate statistical measures on them. I will introduce the visual mapping and in a short demo, I will demonstrate the core features of our system.

The increased interest in neural networks and their common use imposes the question about the ‘How ?’ - How do they work? What do they capture?. I will present a visual tool, LSTMVis, that allows investigation of state changes on trained models (LSTMs) and can be one puzzle piece in the quest of white-boxing neuronal networks.

And finally, I will present a project - Forma Fluens - that mixes aspects of art and data science and highlights how the two fields benefit each other.

The projects often resulted in open source tools which are used 1) by domain experts to help them to generate hypothesis as a first step towards insight or 2) to inform the general audience about insights.

Hendrik Strobelt creates visual analysis tools and explores visual encodings to​ ​facilitate discovery. His work at IBM Research AI focusses on explainable​ ​neural networks. Before that, he was Postdoc at Harvard SEAS and NYU Tandon.​ ​He received his PhD from University of Konstanz, Germany - advised by Oliver Deussen.​ ​He graduated with a MSc (Diplom) from Technische Universitaet Dresden, Germany.

​For more information ​at:

Host: Professor James Tompkin