Ensemble Learning for Spiking Neural Netwoks

The area of Spiking Neural Networks (SNNs) has emerged as a result of recent physiological exper-
iments which show that in many areas of the nervous system, information is encoded in the timing
of the individual spikes and not in the ring rates of neurons which is encoded by Articial Neural
Networks (ANNs) (Ponulak and Kasinski, 2011). Due to the heavy influence of biological concepts
that are not well understood and the vast number of approaches applied when designing neural
models, SNNs have been unable to achieve performance comparable to artificial models, even on
simple datasets.
Ensemble learning has been a long investigated topic in the context of ANNs due to the properties it
provides, which are independent of individual model characteristics. From a theoretical perspective
ensemble classiers are guaranteed to achieve better accuracy when compared to individual models,
as demonstrated by the Ambiguity Decomposition (Krogh, 1995). Furthermore, through the use
of Bias-Variance-Covariance Decomposition, additional insight can be gained into the causes of
classication error. A learning routine that makes use of all these advantages is Negative Correlation
(Brown et al., 2005).
After an overview of the literature of SNNs, it has become apparent that no ensemble learning
approaches have been applied. Thus, we believe that through the use of Negative Correlation
learning, we would not only improve the performance of existing models, but also provide additional
insight into the properties of the models used and the ffects of diversity on their performance.