PhD position: Structured Dictionary Learning for Brain Computer Interfaces
In Brain Computer Interfaces (BCI), calibration of the system is necessary, due to the high variability occuring in brain signals. One can identify at least three types of variability to deal with:
-
across subjects;
-
for a given subject, across sessions;
-
within a single session, over time.
From a machine learning point of view, this high variability means that models trained on data from a single session and subject will generalise poorly to different sessions and subjects, and may not perform well. This is the reason why calibration is generally performed at the beginning of each session.
This PhD aims to provide novel auto-calibration strategies, in order for a BCI user to immediately receive feedback. This feedback should originate from generic features and classifiers, which are subsequently tuned to the user. The features and their classifiers will need to take into account the various types of variability present in the EEG. For this, we will design specific dictionaries for the features, and use generic classification methods.
A way to do this is to design a structured dictionary, which will contain the relevant features for BCI. The idea is to construct sets of discriminative features that are robust to noise, and to the modifications that occur across sessions, and across subjects. This dictionary will be learned from the data, and constrained to have some structure. We propose to use a dictionary structure inherited from prior assumptions on the variability in the data. Variability priors may concern how variability occurs across the different dimensions of the data (space / time / frequency), or the hierarchical nature of the variability. Once the dictionary is learned, features are estimated by solving a sparse coding problem.
Contrarily to classical Dictionary Learning, here the purpose of the dictionary is not to obtain a sparse representation of the data, but to obtain sparse discriminative features. Our Structured Dictionary Learning methodology will therefore incorporate the classification strategy.
Context:
This PhD will be cosupervised by Maureen Clerc (Inria Sophia Antipolis) and Alain Rakotomamonjy (University of Rouen).
It will take place at Inria Sophia Antipolis, within the Athena project-team.
Required skills:
A strong background in machine learning and signal processing is required. Fluency in English, skills in computer programming (notably C++) as well as ability and motivation to work in a multidisciplinary environment are necessary.
