Valentin
Jacquin

I have always thought of the brain as a computer, using biology as its architecture. That is why I did a double degree at UVSQ, combining computer science and biology. I wanted to have both sides to use them together in my future careers.

Since then, I graduated from a Master’s in Cognitive Science (Cog-SUP, Sorbonne Université and Université Paris Cité), on the computational neuroscience and AI track. My thesis was done at the Group of Neural Theory at ENS Paris, where I modeled how the loss of nicotinic receptor signaling in Alzheimer’s changes the firing dynamics of a cortical population. I wrote the model from scratch in Python, and most of the work was in the fitting. I developed a loss based optimization in order to find parameters that produce a bistable network, matching the real recordings in both the high and the low firing rate states. Before that, during my M1, I built a transformer model at Institut Pasteur to predict music sequences and used its prediction error as a measure of musical surprise, comparing it to neural data.

Alongside that, I have been working with the CERVO centre in Quebec on a speech pipeline for schizophrenia research, and this is the work I own end to end. I designed the pipeline, benchmarked several architectures across a thousand audio files, and run every model myself on my own GPU workstation. I also host a small web server so the team can access the results directly, replacing a step they were doing entirely by hand until then. The research and the implementation are mine, with advice from Antoine Dufour (Predis) and Evann Rabeau along the way. I was the one doing the technical implementation, and that is exactly why it is the experience I have learned the most from.

This is also what interests me most about AI in science. The point is not a better benchmark score, it is closing the loop between a hypothesis and the experiment that tests it. That translates into giving AI agents hands: a real lab they can act on, so an agent proposes something, tests it for real, sees the result and improves from it, instead of stopping at a prediction. Automated labs, tools that remove manual steps, systems that let a scientist run ten experiments where they used to run one. I want to be part of building the new tools for scientists.

Outside of work I do sport, mostly triathlon. It is how I experience life through my body rather than through my head, and I need both.