Laboratory : BMBI - UTC
Place : Compiègne, France
Key words : machine learning, representation learning, deep learning, sparsity, classification
Starting date : October 2021
The performance of any classification system heavily dependents on finding a good and suitable feature representation space where observations from different classes are well separated. Unfortunately, finding this proper representation is a challenging problem which has taken a huge interest in several communities. Indeed, much of the actual effort in deploying machine learning algorithms goes into the design of preprocessing pipelines and data transformations that result in a representation of the data that can support effective machine learning. It has been shown that learning effective representations of the data helps to extract useful information when building classifiers or other predictors.
The representative methods range from the early-staged hand-crafted feature design (e.g. SIFT, LBP, HoG, etc.), to the feature extraction (e.g. PCA, LDA, LLE, etc.) and feature selection (e.g. sparsity-based and submodularity-based methods) in the past two decades, until the recent deep neural networks (e.g. CNN, RNN, etc.) achieving the state-of-the-art performances.
The goal of the proposed thesis is to introduce novel, efficient and interpretable data-driven representation learning techniques. We will particularly interest to sparse, hybrid (sparse and dense) as well as deep representations with a focus on the interpretability for the target application. Indeed, interpretability is of extreme importance in order to describe the internals of a system in a way that is understandable to humans. These explanations are important to ensure algorithm fairness, identify potential bias/problems in the training data and robustness against uncertainties, as well as to ensure that the algorithms perform as expected.
For the targeted application, we will have a particular focus on multidimensional biomedical data and more especially simulated and recorded High-Density Surface Electromyogram (HD-sEMG) measuring the electrical neuromuscular activity in a non-invasive way. Possible applications include and not limited to, muscle aging evaluation and control of prosthetic devices for individuals with amputations or congenitally deficient limbs.
BMBI UMR CNRS 7338 BioMécanique et BioIngénierie. Université de Technologie de Compiègne (UTC) Centre de Recherches de Royallieu Rue Personne de Roberval - 60200 Compiègne
(c) GdR 720 ISIS - CNRS - 2011-2020.