Tailored Time-Frequency Features for Robust Classification of Electrophysiological Correlates of Human Memory Retrieval

Memory is a ubiquitous part of everyday life, and advancing our understanding of memory will not only inform basic science but be of important applied and clinical relevance. Currently, there is little direct evidence of the temporal dynamics of memory reactivation and the proposed research will provide this missing information by developing robust tools to capture the real-time activation of memories in high temporal resolution multi-channel recordings of brain activity. We will study the electrical brain signals, EEG, which are known to have the required time-resolution, although the noise level is very high. It is therefore important to extract relevant and robust features to capture the time- and frequency variations.

Studying a time-frequency (TF) image, e.g. a spectrogram, often give excellent information and visual analysis based on this view has the potential to be very efficient, as an expert knows what to look for and which parts to ignore. The proposed project intends to develop new tools for TF image feature extraction, where the focus will be on non-stationary multi-component signals. These novel methods will allow us to, in previously unprecedented ways, especially address key issues in the cognitive neuroscience of memory.

In this project we will focus on a special type of signal model, for which we can find several interesting application areas for classification, and where most existing methods will fail. The signal model is multi-component, meaning that several components (atoms) are detectable in the TF image. The signals are allowed to have a component-wise stochastic variation, in time, frequency as well as amplitude and signals in different classes might share TF overlapping components of strong power where important between-class differences are located to certain deviations in time, frequency or amplitude of possibly weaker components. Such signals call for sensitive and tailored methods.

Research group

PI:Maria Sandsten
Mathematical Statistics, Centre for Mathematical Sciences, Lund University
Rachele Anderson
Mathematical Statistics, Centre for Mathematical Sciences, Lund University
Mikael Johansson, Inês Bramão
Department of Psychology, Lund University