Explainable Spiking Neural Networks for Model Developers Using Interactive Treemapping : A Visual Analytics Design Study in Neuromorphic Computing

Sammanfattning

This thesis addresses the challenge of explaining Spiking Neural Networks (SNNs) to model developers using interactive treemapping supported by secondary visualizations. SNNs, inspired by the brain’s neural mechanisms, accumulate signals and transmit so-called spikes when the signal reaches a certain threshold. Spikes recorded over time are termed spike trains, which this work primarily relies on. SNNs offer significant advantages in terms of computational efficiency and power consumption compared to traditional artificial neural networks (ANNs), making them an exciting avenue of research. However, their complex nature poses challenges for interpretability, which is critical for debugging, improving models, and ensuring trustworthiness in their predictions. The field of explaining ANNs to make them interpretable is called eXplainable Artificial Intelligence (XAI). XAI is a large field but lacks extensive SNN- specific research, specifically in exploring the hierarchical structures of spike trains. The primary objective of this study is to design a visual analytics tool that facilitates the understanding of SNNs. A design study was conducted in close collaboration with SNN developers. After their needs were identified, the design was adapted to them and the research field gaps. The tool leverages circular treemapping, a method for visualizing hierarchical data, to present the structure and behavior of an SNN interactively. The tool also uses secondary visualizations to provide additional information about selected objects. The resulting system, Cpikes, allows the user to detect inactive or overly active neurons, neurons whose spike train is dissimilar to the target neuron’s spike train, and neuron spike train similarity in general. Several paths for future work were identified, such as enabling easier distinguishing of a neuron’s layer, model comparison, and exploration of other spike train distance metrics. Cpikes is a promising XAI tool that can further develop to be a valuable aid for SNN experts in their process of refining models, potentially leading to advancements on the journey of realizing the full capability of SNNs.