Forum Numerica - Laurent PERRINET (CNRS/Aix-Marseille university): Modelling memory to understand the brain

Modelling working memory ---the ability to store and recall precise temporal patterns--- remains an open challenge in computational neuroscience. While recent advances evolved from Long Short-Term Memory networks to transformer architectures, which excel at linking long-range associations, these models deviate from biological plausibility: they are massive and energy-intensive. I will present our recent work on recurrent spiking neural networks (SNNs), inspired by the structure of information processing in biological systems. Based on the dynamics of realistic yet simple single-neuron models, our approach sheds light on fundamental neural phenomena, such as pattern formation in neural populations, the organization of high-dimensional brain activity, and traveling waves in sensory areas. It also demonstrates practical applications, including the production of gait patterns or the storage of textual information. This work aims to bridge the gap between biologically grounded models and scalable computational solutions, paving the way for more efficient and interpretable AI systems.

Laurent Perrinet is a researcher who studies how the brain performs ‘calculations’. His aim is to recreate large-scale neural networks using artificial intelligence tools. In particular, this could lead to a better understanding of what happens between two eye blinks: how the brain perceives an image, processes it and decides on an action such as eye movement. Laurent Perrinet is currently conducting his research at the Institut de Neurosciences de la Timone (France), a joint research unit (CNRS / Aix-Marseille University).