Cs seminar Niklas Heim

CS seminar - NIKLAS HEIM


Title: nomaly Detection In Chaotic Time Series
- A Reservoir Computing Approachn

Room: E-205

When: [Thu, Oct 4, 2:30 pm] 

This talk will present a method that is able to automatically detect anomalies
in non-linear, dynamical systems. We will assume no prior information about the
physics that govern these dynamics, so there is no knowledge about the kind of
anomaly that we are looking for. This is motivated by the large amounts of
output that state of the art, eddy-resolving ocean models produce. These large
datasets might contain unknown physical behaviour, such as the recently
discovered anomaly of the Kuroshio (an ocean current on the coast of Japan,
which oscillates between two distinct states).

The detection problem is approached by trying to define what is normal, so that
everything that looks significantly different from this norm can be treated as
anomalous. The norm is found by predicting the future evolution of a system
that has been observed for a certain amount of time. This is not a trivial
task, because non-linear systems can exhibit chaotic behaviour which makes
their prediction notoriously hard. However, it can be solved by employing a
so-called "Echo State Network", which belongs to the class of reservoir
computing methods. Once the prediction is extracted from the network, it can be
compared to the true values of the dataset and where they deviate
significantly, a potential anomaly is found.

A bit about Niklas Heim:
Niklas is former student of Computational Physics at the Niels Bohr Institute
in Copenhagen, who will present the results of his Master's Thesis. He is from
Germany, has obtained his Bachelor's degree in Physics from TU Darmstadt, while
also studying at CVUT for one year, and is now hoping to come back to Prague to
join AIC.