Modélisation et conception par approche synchrone d'architectures neuronales hybrides biologique-artificiel
Rasamuel, Marino · HAL (Le Centre pour la Communication Scientifique Directe) · 2023
As Artificial Neural Networks (ANNs) continue to advance in fields like machine learning, robotics, autonomous vehicles, and healthcare diagnostics, an application domain is gaining attraction in both academic and industrial sectors : Neurobiohybridization. This domain seeks to establish connections between artificial and biological neurons with the goal of understanding and potentially repairing or replacing lost brain functions due to disease or accidents. In pursuit of this, the development of biologically inspired artificial neural networks, often referred to as Spiking Neural Networks (SNNs), is essential to enhance compatibility between artificial and biological neural systems. Our thesis fits into this context by using the synchronous approach to model, implement, and simulate bio-inspired and biomimetic SNNs. Leveraging model checkers, that allow to prove or extract properties in systems in formal manner, our aim is to gain a more comprehensive understanding of biological behaviors in the future. For the first time in this context, we utilize the Light Esterel language to achieve our objectives. We demonstrate its potential in implementing neural models, initiating a library of models for exploring different types of SNNs. Throughout this thesis, we developed an entire framework based on LE in order to model, simulate and implement various SNN models. To address neurobiohybridization experiments, we developped our own hardware architecture, SynchNN, capable of executing recurrent SNNs in real-time using our library of models. This framework we developed is completed with an on-going simulation framework aiming to conduct neurobiohybrid experiments in the future.