Clustering basé sur la complexité de Kolmogorov pour la classification de signaux dans le contexte du modèle neuronal d'Izhikevich
Year: 2024
Type: dissertation
Abstract: Spiking neural networks have gained significant attention in recent years due to their energy efficiency. However, major performance challenges still need to be addressed. This study aimed to develop a simple neural network architecture for recognizing and classifying digits without a learning phase, with the goal of performing classification tasks using minimal data and energy. A non-supervised DL classification method using Kolmogorov complexity optimized Kohonen maps and spiking neurons has been proposed. The evaluation demonstrated that Kolmogorov clusters enhanced signal class discrimination. A proposed weighted DTW algorithm showed a slight advantage in performance, though overall scores were low, with a recognition accuracy of about 11%. Confusion matrices revealed persistent class confusions, indicating challenges in accurately differentiating neuronal signal classes. These results suggest the need for further refinement of these techniques. This study opens the door for several promising research directions, including the development of new similarity measures and optimization of clustering algorithms. The results highlight the challenges and opportunities related to the full use of neural networks as they are complex dynamic systems to study.
Source: HAL
Author: Lasserre, Ludivine
Institution: Université des Antilles
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7135157170