SCCS Colloquium - Aug 28, 2019

From Sccswiki
Jump to navigation Jump to search
Date: August 28, 2019
Room: 02.07.023
Time: 15:00 - 16:00

Hung Phu Nguyen: Seismic Hazard Map

This is a talk for an Application Project in the Data Engineering and Analytics program. Advisor: Carsten Uphoff.

Seismic simulation data can be used to visualize seismic hazards. The project aims at generating hazard maps out of the data by mapping the data to a geographical map, and showing statistics of the simulation, using established measures from the geoscientific community. The project should make use of OpenStreetMap data. The output should be a slippy map displaying simulation data in a layer above the layer of geographical data retrieved from OpenStreetMap. It is required that the application is able to correctly display data sources coming from different geographical projections. The project will need to deal with multiple types of data and a large amount of data (however, the data might not need to be processed on the fly but only transformed and stored statically): Inputs are simulation data files which may consist of raw data sizes in the Terabyte range, and the project is required to transform the data into another format supported by some framework to display on the map.

Keywords: seismic hazards, openstreetmap, visualization

Language: English

Yelysei Bondarenko: Quantum Variational Learning of Generative Neural Networks

This is a Master's thesis submission talk. Yelysei is advised by Prof. Hans-Joachim Bungartz.

Generative neural networks are powerful probabilistic unsupervised learning algorithms. Currently, they are one of the most promising approaches towards analyzing and understanding large corpora of unlabeled data. Recent advances in quantum computing suggest that we can use quantum devices for sampling from certain complex distributions more efficiently than it is possible on a classical computer.

Inspired by the success of classical Boltzmann machines, we introduce a novel hybrid quantum-classical learning algorithm that employs the quantum approximate optimization algorithm (QAOA) as a subroutine in order to approximately sample from Boltzmann distribution, and potentially to speed up the training of Boltzmann machines.

We focus on fully-visible fully-connected case and evaluate the proposed method on toy dataset and compare the performance against classical counterparts. The proposed method is simulated on a classical computer using our TensorFlow + Cirq implementation.

Keywords: unsupervised learning, generative models, Boltzmann machines, gate-model quantum computing, quantum variational algorithms

Language: English