We have decided to discontinue the publication of preprints on our preprint server as of 1 March 2024. The publication culture within mathematics has changed so much due to the rise of repositories such as ArXiV (www.arxiv.org) that we are encouraging all institute members to make their preprints available there. An institute's repository in its previous form is, therefore, unnecessary. The preprints published to date will remain available here, but we will not add any new preprints here.
MiS Preprint
96/2019
Training Delays in Spiking Neural Networks
Laura State
Abstract
Artificial Neural Networks (ANNs) are a state-of-the-art technique in machine learning, showing high performance in many different tasks. However, their demand for computational resources is high, both during training and testing. An alternative framework is provided by Spiking Neural Networks (SNNs), a model that is closely inspired by biological networks. The energy consumption of SNNs is small, however, their performance lies below that of ANNs. The main reason for this gap is the much harder training of SNNs. In this thesis, we propose a new supervised framework for training SNNs. Inspired by research in theoretical neuroscience that highlights the importance of temporal codes, we introduce a delay parameter. We propose two different training approaches: a transformation to the complex domain combined with a linear regression and a standard gradient descent. We evaluate our training framework on two different classification tasks, based on a synthetic dataset and the MNIST dataset of handwritten digits. A single-layer network trained by both approaches is able to perform the given classifications tasks. Our supervised framework provides a new approach for training SNNs and can be used to optimize the training of neuromorphic chips.
