Interactive Analysis of Spiking Neural Networks Simulation Traces

Résumé

Neuromorphic architectures are promising approaches to significantly reduce the energy consumption for tomorrow’s computers and the post-Moore era. The brain function is the inspiration behind this architecture, consisting of spiking artificial neural networks. Due to low energy consumption, deploying such architectures is useful in many applications, especially those with energy-limited constraints. Furthermore, using this architecture, we can process a large quantity of data and provide the computation power needed for machine learning tasks. Neuromorphic architectures consist of spiking artificial neural networks inspired by the brain functionalities with many open questions. This situation impacts the implementation of artificial spiking neural networks and their performance compared to conventional neural networks. Moreover, in SNN, many questions are still debatable, like how the learning is happening and what learning rule is the most suitable, memory location in such networks and how it works, and how the network encodes the information using spikes. Such neuroscience-related questions prevent spiking neural networks from performing like the conventional ones. Therefore, to better understand the different phenomena in SNN, we need to analyze the internal network activity during the simulation. The network activity contains the spikes, neurons, and synapses states activity. When simulating a large network that takes time and resources to finish, we generate a large simulation trace that is challenging to analyze due to its size and spatio-temporal aspect, which we can study at several scales. This manuscript aims to study the visual analysis of spiking neural networks by visualizing the collected trace from a simulation. The primary objective is to better understand the different network phenomena and improve the network using visual analysis. The first contribution is the study of the visualization techniques in SNN simulators. This study from the technical and visualization aspect of the simulators shed light on the diversity of the used technologies. Furthermore, this study also shows the similarity of the visualization techniques provided by the simulators. At the end of this study, we concluded that we need more dedicated tools to analyze than what simulators provide for visual analysis. Next, we developed VS2N (Visualization tool for Spiking Neural Networks). A web-based tool for post-mortem interactive dynamic visualization and analysis of spiking neural networks. The novelty of VS2N compared to the existing visual analysis tools can be summarized in four points: modular nature, simulator-independent, scalability, and dynamic analytics. In addition, VS2N provides the possibility to walk in time with the evolution of the network during activity, which is not possible using the existing tools. This feature is significant when the network evolution is over hours of activity, which is the case in spiking neural networks. Finally, we proposed a novel approach to compress a spiking neural network based on the visual analysis conducted on SNN. This dynamic compression approach concerns the synapses in the network by providing two formulas to calculate the dynamic threshold, which changes based on the compression status, instead of having a static threshold which is the case in the existing works. This approach can maintain or improve the network accuracy compared to the non-compressed network while compressing up to 80%.

Citer ce document

Accès au document

Ce document est hébergé sur son dépôt institutionnel d'origine.

Auteur(s)

Statistiques

Consultations : 3

Téléchargements : 0