For our next talk, in the Machine Learning seminar series, we will be hearing from Professor Hernando Ombao, Professor of Statistics, King Abdullah University of Science and Technology. We’re delighted to host Hernando in what promises to be a great talk!

Talk title: Time Series Models with Applications to Brain Connectivity

Date: Wednesday 17 September
Time: 1:00 – 2:00 pm
Location: BDI/OxPop Seminar room 0

Abstract: This talk will cover some of the recent work on characterizing brain connectivity in a network. First, we present spectral transfer entropy (STE) which is a novel methodology that captures cross-channel information transfer in the frequency domain. STE quantifies both the magnitude and direction of information flow from a band-specific oscillation of one channel to another band-specific oscillation of another channel. As a dependence metric, the advantage of STE is that it uses joint the joint probability density of the oscillations. Another novel aspect of this novel contribution is a simple yet efficient method for estimating STE using vine copula theory.

In the second part of the talk, we discuss the problem of modeling dependence between two cortical regions of the brain where each region contains multiple signal recordings from several channels. An exploratory tool for studying the dependence between two random vectors is via canonical correlation analysis. Mitigating the limitations of current approaches (i.e., sensitivity to outliers and inability to capture non-linear dependence), we develop a robust method, Kendall’s tau-based canonical coherence (KenCoh), to learn connectivity structure among neuronal signals filtered at given frequency bands.

The third part of the talk is about Granger Causality (GC) in a brain functional network. The classical approach is built on vector autoregressive models (VAR) which are simple and easy to fit, but have limited practical application because of their inherent inability to capture more complex (e.g., non-linear) associations. Here, we develop an approach that exploit the functional approximation power of deep neural networks (DNNs) for GC. We present a novel paradigm for investigating the learned GC from a single neural network that joint models all components of multivariate time series data. We propose to uncover the learned GC structure by comparing the model uncertainty or distribution of the errors for the network that uses the past of everything as compared to a network where a specific time series component is dropped from the model.

The fourth part of the talk is on modeling lead-lag dependence between a pair of regions (or channels) when the time lag varies across frequency oscillations (e.g., millisecond oscilatory processes may be perfectly contemporaneous, but lower frequency oscillations may exhibit lags). We will present the generalized Cramer representation and propose a procedure for estimating the frequency phase plots.

The work presented here are in collaboration with Paolo Redondo, Mara Talento, Malik Shahid, Maurizio Fillipone, Sipan Aslan and Raphael Huser, all from KAUST.

Bio:
Hernando Ombao is a professor in the Statistics Program and the principal investigator of the Biostatistics Group at KAUST. His research focuses on developing time series models and novel data science methods for analyzing high-dimensional complex biological processes. He leads a group of researchers specializing in spectral and time-series analysis, functional data analysis, state-space models, and signal processing for brain signals and images. His group collaborates closely with neuroscientists to model the associations between neurophysiology, cognition and animal behavior.
————————————————————————————————————————
All members of the University are welcome to join, please let reception at BDI know you’re here for the seminar and sign-in. We hope you can join us!
We encourage in-person attendance where possible. There is time for discussion over, tea, coffee and pastries after the talks.

Hybrid Option:
Please note that these meetings are closed meetings and only open to members of the University of Oxford to encourage sharing of new and unpublished data. Please respect our speakers and do not share the link with anyone outside of the university.

Join the meeting now
Meeting ID: 320 434 947 113 3
Passcode: aU25r4NR