Recurrent SNNs to solve the XOR problem
Keywords
Loading...
Authors
Issue Date
2022-06-20
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Spiking neural networks (SNNs), the biologically plausible successor of artificial
neural networks are growing in popularity. A lot of work focuses on improving
learning algorithms, figuring out ways to encode information in spikes as efficiently
as possible. This work shows that recurrent connections can improve the
accuracy of temporally encoded SNNs using spike-timing dependent plasticity
(STDP) modulated by a global reward signal on the task of learning the inputoutput
mapping of the XOR gate. Furthermore, we show that the additional
recurrent connections might lead to a decrease in accuracy for rate encoded networks.
Lastly, we show that problems with rate encoding, reward-modulated
STDP and the XOR problem can be overcome if the spike rate is reduced by
e.g. resetting the STDP traces and the membrane potentials of neurons after
output spikes.
Description
Citation
Supervisor
Faculty
Faculteit der Sociale Wetenschappen