2024/11/27 - BrainHack Marseille 2024
Taking part of the organisation of the 2024 version of the BrainHack Marseille !
2024/10/01 - Begining of the PhD...!
Fellowship from the doctoral school “Ecole Doctorale Sciences de la Vie et de la Santé, Aix Marseille Université”. Supervised by Laurent U. Perrinet and Sophie Denève.Project description :
Recent technological advances in neurobiology have paved the way for the recording of very large populations of neurons at the resolution of action potentials (spikes). They shed new light on the structure of neuronal activity, and in particular on the rich spatio- temporal structure of neuronal information. Current analysis methods are not yet adapted to this level of precision, and the aim of this thesis project is to develop new methods for analyzing neurobiological data that take into account some fundamental principles derived from computational neuroscience. The main objective is to bridge the gap between two classical approaches, the encoding and decoding of neural activity, in a self-consistent way, i.e. by achieving coherence between the way neurons encode information and the way it can be decoded.
To achieve this, the project will draw on a theoretical approach developed in our group, which allows us to formalize this consistency by deducing a measure of the efficiency of processing spikes. The analysis algorithm will be optimized by minimizing a cost function reflecting these principles. Existing encoding and decoding models, which we have already validated on neurophysiological data, will then be combined using this approach. By studying the role of principles such as the energy efficiency of neural processing and the consideration of physiological constraints, we will then be able to deduce the role of each of these principles in neural information by measuring changes in processing efficiency.
Ultimately, the aim is to develop new analysis methods enabling predictive links to be established between recorded multidimensional neural activity and functional hypotheses such as object detection or spatial navigation. This method will also enable us to test various hypotheses on the role of temporal precision in information processing. The risks associated with the innovative aspect of the project will be mitigated by the supervisors’ expertise in machine learning, neural modeling and computational neuroscience.
2024/06/25 - Work on Titanic Dataset !
Using PyTorch to predict the future of the Titanic's passengers based on the information rescued from the dataset. This notebook is composed by 4 parts :- Data importation and preprocessing of the dataset
- Model preparation (construction and training function)
- Model training
- Model optimisation with optuna