Patricia-Reynaud Bouret - CNRS Silver medal in mathematics, 2021 - 3IA Chair “Modeling and estimating learning”
#Statistics #learning algorithms #neuroscience cognition
Understand how animals and humans are learning tasks
We want to understand how animals and humans are learning tasks. To do so, we look at classical and less classical learning algorithms such as reinforcement learning, adversarial bandits and activity-based credit assignment (ACA). The aim is to to fit these algorithms on behavioral data and to match the different behaviors with neuronal network activities that have been recorded at the same time.
A combination of several type of AI
There are various type of artificial intelligence involved in the project: Lasso techniques and bootstrap to match neuronal activity with behavior as well as learning algorithms to fit the learning dynamics
Analysis of spikes trains and functional connectivity
We hope to discover the path of the mnesic trace inside the brain: what are the cerebral regions involved in learning? when are they involved ? at which step of the learning are they important?
A collaboration with actors of University Côte d’Azur ecosystem in the Nice / Sophia Antipolis region
Most of my collaborators are there belong to the newly found component of Université Côte d’Azur: NeuroMod, an interdisciplinary institute on modeling in neuroscience and cognition, that I lead. With more than 200 members in our community, this is precisely the fertile environment that we need to develop this very interdisciplinary project.
I collaborate mainly with Ingrid Bethus, neuroscientist at IPMC (UCA), Francesca Sargolini, neuroscientist at LNC (Marseille), Alexandre Muzy computer scientist at I3S (UCA) specialized in ACA and Luc Lehericy theoretical statistician at LJAD (UCA). Let us not forget our Phd students : G. Mezzadri (who works in math-psycho under the supervision of F. Mathy (BCL, UCA) , T. Laloë (LJAD, UCA) and myself) and A. James (who works in computer science -neuro under the supervision of A. Muzy and I. Bethus).
Success story!
We just found out that we can infer which path a rat is taking in a maze, based on the neuronal recording of a simple path in a maze (whose duration is typically less than 1s). The method is based on functional connectivity with spike trains. This leads us to the notion of encoding power of a cerebral region which depends on the cerebral region as well as the stage of learning.
List of the main publications:
-
Phi, Tien Cuong; Muzy, Alexandre; Reynaud-Bouret, Patricia Event scheduling algorithms with Kalikow decomposition for simulating potentially infinite neuronal networks, SN Computer Science, 1(35) (2020).
-
Ost, Guilherme; Reynaud-Bouret, Patricia Sparse space-time models: concentration inequalities and Lasso, Annales de l'IHP, Probabilités et Statistiques, (2019).
- Lambert, Régis; Tuleau-Malot, Christine; Bessaih, Thomas; Rivoirard, Vincent; Bouret, Yann; Leresche, Nathalie; Reynaud-Bouret, Patricia Reconstructing the functional connectivity of multiple spike trains using Hawkes models, Journal of Neuroscience Methods, 297, 9-21 (2018) .
This figure shows on few recordings of a rat in a maze how the neuronal activity might help to predict the taken path. Depending on the experiment, the number of taken paths varies and the black triangle shows the probability to pick the correct path uniformly at random. Depending on the model we use, the prediction gets better. The best one obtained in 85% of the cases was the one called “Hawkes model” which uses the functional connectivity between neurons to model their activity. It means that the neural network itself in its dependence helps us to decode the behavior of the rat.
