AI for Time Series

every two weeks, Leuven & online
seminar series
VAIA, Flanders AI Research and KU Leuven STADIUS

More often than not, measured data or observations come in the form of time series: scalar or vector sequences of data points indexed in time order. Time series show up in many fields of society, science and engineering. Due to the availability of massive amounts of data and increased computing power, the analysis of time series has become increasingly important.

Time series analysis has various purposes, which can be mainly categorized into simulation, prediction, clustering, and anomaly detection. Another vast area is model-based control, where dynamical models are obtained from time series.

Several research groups in the Flanders AI Research Program (FLAIR) conduct world-class research on time series, both in the development of algorithms and tools, as in a wide area of application fields. In a recent poll in the Flanders AI community, ‘time series’ came up as the most wanted topic for future workshops or courses. With this seminar series we bring together researchers that are interested in, or are conducting research related to time series. We offer a varied program of national and international speakers.

AI for Time Series Seminars

This talk explores opportunities in combining insights from systems theory with tools from machine learning, for data driven modelling of dynamic systems.

John Lataire discusses the data-driven modelling of Linear Time-Invariant (LTI) systems with Gaussian Processes (GP) regression. First, GP regression in general, and Time domain and frequency domain expressions of LTI systems are reviewed. Then, relevant system-specific properties, s.a. causality and stability are encoded in kernels, used as prior knowledge for estimation purposes.

Care is given to visual interpretations of this prior knowledge in the spectral domain. Finally, recent results are shared on the estimation of the more challenging situation, where the systems are lightly damped. The combination of the non-parametric local rational model (LRM) estimator with the GP regression approach is proposed.

After an introduction to the field, we will discuss how the marriage of power systems engineering and artificial intelligence can aid to improve the visibility on the low voltage distribution grid. This enables optimized use of the presently installed assets and helps to limit future investments, so the distribution grid becomes an enabler of the energy transition, rather than a costly bottleneck.

Currently, sports is an incredibly data rich domain as it is possible to collect massive amounts of data from both training sessions and matches. Typically, this data comes in the form of time series such as sensor data (e.g., accelerations, heart rate, GPS, etc.), event stream data, and optical tracking data. The availability of this data has driven an explosion of interest in the automated analysis of sports. The goal of this talk is to provide an overview of this area with illustrative examples arising out of work done in my research group.

I will present work being conducted in the Computational Creativity Lab at the VUB which explores the use of statistical models to build cognitive representations of sequence data. Sequence modelling is a central part of cognitive systems. The notion of sequence can be thought of as an abstraction of time series used to account for cognitive flexibility in the perception of time. The goal of our work is to bridge the gap between unsupervised learning and knowledge-based reasoning in cognitive systems.

A deep neural network model typically is learned solely from data in the form of input-output pairs ignoring domain knowledge that additionally might be available. When domain knowledge is injected to serve as a learning guidance it is expected that the need for annotated data is relaxed and training is less dependent on the initialization of the model parameters. Domain knowledge can be provided in many different types and forms. In this seminar, domain knowledge is assumed to be given as inequality constraints which appear naturally in practice. Compared to other neuro-symbolic approaches the discussed method is also able to incorporate non-linear inequality constraints and does not require to first transform the constraints into some ad-hoc term that can be added to the learning (optimisation) objective.

The definition of adequate metrics between objects to be compared is at the core of many machine learning methods (e.g., nearest neighbors, kernel machines, etc.). When complex objects are involved, such metrics have to be carefully designed in order to leverage on desired notions of similarity.
This talk covers my works related to the definition of new metrics for structured data such as time series or graphs.

Practical

  • every two weeks,
    starting on 28 October 2021
    14.30-15.30h
  • Location: Campus Arenberg, Leuven
    + online streaming
  • Language: English
  • Contact: Philippe Dreesen & Katrien De Cock
  • Target audience: everyone interested in research on AI and time series

Registration

  • Price: free
  • Registration is not necessary, you can simply join the seminar.
    However, if you would like to be reminded of this seminar, fill in your contact details below and you will receive a notification on the day of the seminar.

Ready to get started

the link to the online seminar environment will be put here shortly

Remind me of the
AI for Time Series Seminars

Register here to receive reminders of the AI for Time Series Seminars.
We will use this information only for this purpose as stipulated in our Privacy Statement & Disclaimer. You can unregister at any time via the unregister link at the bottom of each reminder.