Noise or change? Tuning the speed of learning in an uncertain World
Jérémie Naudé, PI
Equipe Perroy, IGF, Montpellier

Résumé
Learning from mistakes requires telling noise from change: a missed reward may reflect chance, and should then be ignored, or a change in the environment, which calls for new choices. One solution is to learn faster in volatile environments and more slowly in stochastic ones; another is to adjust how much we explore between options. Yet in classical laboratory tasks, making outcomes noisier by manipulating reward probabilities also makes the options harder to tell apart. How we fine-tune our choices in noisy, changing environments has therefore remained unresolved.
We developed a new task in mice that separates volatility, stochasticity and difficulty. Mice raised their learning rate with volatility and lowered it with stochasticity, while their level of exploration remained constant. To account for this, we built a new learning model in which two simple estimates tune the learning rate. Stochasticity is read from the average magnitude of errors. The volatility estimate rises when repeating a choice stops paying off, or when switching starts to pay off. This model best reproduced the choices of mice and generalized across species, as it also accounted for rat decisions under uncertainty. In rats, noradrenaline release in the orbitofrontal cortex tracked the model’s volatility estimate. Chemogenetic inhibition of locus coeruleus inputs to this cortex reproduced the deficits predicted by removing the volatility component of the model.
We thus suggest that adapting to noisy, changing environments relies on a simple rule that sets the learning rate by weighing volatility against stochasticity, the former being signaled by noradrenaline in the orbitofrontal cortex.

Invité par David Robbe
Mardi 20 octobre 2026 à 11h – Salle de conférence de l’Inmed