Maren Eberle: Neuronal Excitability Classes in Spiking Neural Networks

BCCN Berlin / Technische Universität Berlin

Abstract

Spiking Neural Networks (SNNs) emulate more biological detail than other artificial neural networks to achieve higher efficiency. The units of the network are described by mathematical single-neuron models. In SNNs, heterogeneity in neuron-intrinsic parameters such as time constants improves performance on tasks with a rich temporal structure (Perez-Nieves et al., 2021). In biological neural networks, heterogeneity is universal for properties of neurons and synapses. Specifically, neurons have different dynamical types, called neuronal excitability classes, describing their computational properties. Excitability classes affect the activity in whole networks, for example, its synchrony and encoding properties. In this work, the effect of emulating heterogeneous neuronal excitability classes in recurrent SNNs on performance and network connectivity was investigated. Training approaches optimizing only network weights, and co-training the excitability class for each neuron with the weights were compared across SNNs with leaky integrate-and-fire (LIF) and quadratic integrate-and-fire (QIF) neurons. Unlike the LIF, the QIF neuron model can exhibit more than one excitability class. Using a supervised classification setting with surrogate gradients, networks were trained on three data sets with increasing temporal structure. Heterogeneous excitability classes were successfully learned. When controlling for confounding effects of optimizing excitability classes during training, neuron dynamics tended to accumulate close to a switch between excitability classes. A relationship between the distribution of excitability classes and connectivity patterns between neurons in SNNs is indicated. Future work should explore which training algorithms and neuron models can show the effects of heterogeneous excitability classes. This can help shed light on the role of neuronal excitability classes in biological networks.

 

Additional information:

Master thesis defense

 

Organized by:

Prof. Dr. Susanne Schreiber & Prof. Dr. Kerstin Ritter