A Preliminary Analysis on Software Frameworks for the Development of Spiking Neural Networks

Resumen

Today, the energy resources used by machine learning methods, especially those based on deep neural networks, pose a serious climate problem. To reduce the energy footprint of these systems, the study and development of energy-efficient neural networks is increasing enormously. Among the different existing proposals, spiking neural networks are a promising alternative to achieve this goal. These methods use activation functions based on sparse binary spikes over time that allow for a significant reduction in energy consumption. However, one of the main drawbacks of these networks is that these activation functions are not derivable, which prevents their direct training in traditional neural network development software. Due to this limitation, the community has developed different training methods for these networks, together with different libraries that implement them. In this paper, different libraries for the development and training of these networks are analysed. Their main features are highlighted with the aim of helping researchers and practitioners in the decision making process regarding the development of spiking neural networks according to their needs.