Research and Teaching
Research: what am I curious about?
Since I began my academic studies, the two rationales that incentivized me to pursue a career in computational neuroscience are understanding how to create intelligence and memory from simple neuronal-like mechanisms, and the endless possibilities for learning and collaboration.
Discrete-event simulation (DEVS)
Discrete-event simulation is a range of formalisms (DEVS) and techniques taking advantage of the time sparsity of state changes in a system to focus computational power to the parts of a system actually undergoing changes. Discrete-event simulation can improve computation speed and reduce energy consumption, in particular when coupled with specialized discrete-event hardware.
Point processes (Hawkes process)
Point processes are a countable, random set of points. They are best described, when it exists, by a conditional intensity function (cif), analogous to an instantaneous rate of occurrence for new points.
The Hawkes process follows the characteristic cif:
The Hawkes point process can be easily used to model neuronal interactions and spiking activity. I have helped design and implement particularly efficient algorithms for the simulation of very large networks of Hawkes processes.
Spiking Neural Network simulation (Spiking neuron models, SNN)
SNN are a type of neural networks for machine learning, typically referred to as the third generation of neural networks. I believe they can be more efficient that neural networks from the first two generations (perceptron-based neural networks and networks based on computational units applying an activation function). I am applying the efficient, discrete-event simulation algorithms developed for a large variety of neuron models in order to improve the performances of SNN.
Perspective
At present I have been one of the lead developers in a graphical interface to the simulators that have been developed during this thesis, and for future projects of the Neuromod Institute, in collaboration with the SED team of INRIA Sophia-Antipolis. The new graphical interface will facilitate the use of our software (SPIKES) and tools (UnitEvent, then RECONSTRUCTION) by experimentalists working on the study of neural coding, and neuronal connectivity. The software will be developped further by new members of the Neuromod team.
I am also one of the lead developers in a more efficient implementation of the UnitEvent package. The method estimates the parameter matrices of the neural network, modeled as a network of Hawkes processes, from the time of spikes of the neurons~\cite{scarella2020}. In collaboration with a research engineer at the LJAD laboratory, Gilles Scarella, I used classical data structures to compute these matrices in a new way, drastically reducing the execution time and the memory imprint. Thanks to the reduced resource consumption, the method shall become usable on large neural networks, thus facilitating the mapping of entire connectome, as well as take full advantage of the latest spike recording chips.
My main project at the moment is to adapt the Hawkes process to STDP, and explore the performances of networks of Hawkes processes in machine learning. I have adapted the Hawkes process, using a simpler version with only Dirac functions for the interactions.
