Student project
Student project
NeuroRestore is a research and innovation center spanning EPFL and the University Hospital of Lausanne (CHUV) that develops and applies medical therapies aimed to restore neurological functions. We integrate implantable neurotechnologies with innovative treatments developed through rigorous preclinical and clinical studies. These developments have led to breakthroughs for the treatment of paraplegia, tetraplegia, Parkinson’s disease, stroke, and traumatic brain injuries. By working with our network of vibrant high-tech start-ups and established medical technology companies, we are committed to validate our medical therapy concepts and see them used every day in rehabilitation clinics worldwide.
Representation learning for motor pattern recognition
This project is devoted to the problem of representation learning for motor pattern recognitionin the context of epidural electrical stimulation (EES). Spinal cord injury disrupts the communicationbetween the brain and the spinal circuits below the lesion that generates and coordinateslimb movements. To restore the movements of paralyzed limbs an external controller is used.It sends wireless commands to the implanted stimulator, which then delivers EES directly overthe spinal cord. The effect of EES significantly varies on the frequency, amplitude, and locationof the stimulation. Even for a simple functional movement such as walking, the calibration isperformed over several weeks by expert engineers.
The present project offers students an opportunity to enhance their proficiency in employingdiverse representation learning methods as well as an ability to work with real-world biomedicaldata. The main task of the project is to analyze the dependency between input electrodeconfiguration and output movement trajectory. We expect improvement in both the modelinterpretability and performance from the usage of tensor decomposition methods [1] as well asusage of biologically-informed prior knowledge in neural models [2]. Previously, researchers usedrepresentation learning methods to analyze rotational neural dynamics [3] as well as kinematicpatterns resulting from EES [4, 5]. These methods can also be employed in this project. Anotherapproach to representation learning we are interested in is distance learning and similaritylearning [6] that can be implied to reveal similar patterns in the model feature space.
Additional information:
What will you learn? Principles of exploratory data analysis and representation learning.
You will also gain experience of working with complex biomedical time series data.
Requirements: Machine learning fundamentals, a good level of Python, experience with
Git. Knowledge of PyTorch and/or passing of Deep learning EE-559 course (or similar)
would be an additional plus.
Supervisors: Oleg Bakhteev oleg.bakhteev@epfl.ch,
Leonid Iosipoi (leonid.iosipoi@epfl.ch), Guillaume Obozinski (guillaume.obozinski@epfl.ch),
Gregory Dumont (gregory.dumont@epfl.ch), Alice Bruel (alice.bruel@epfl.ch).