Implementing Granger Causality using Spiking Neural Networks
Spiking Neural Networks, Granger Causality, Time series.
This work proposes the implementation of the Granger Causality using Spiking Neural Networks (SNNs). The main idea is to detect causality between time series under the assumption that the prediction of one time series is more accurate considering the information of the other. To do so, we intend to build two regression models using SNNs to predict the dependent time series future (our target), using the past samples of it and the past samples of the independent series (our causing series). Spiking Neural Networks are considered the third generation of Artificial Neural Networks, and as its ancestors were developed inspired by the biological behavior of the human brain, using the topology of interconnected neurons. However, unlike its predecessors, they do not activate on each learning cycle, but instead only when the neuron's membrane potential exceeds a determined threshold, making it more energy-efficient on hardware implementations. Being able to capture temporal aspects from data, SNNs had shown to be efficient when working with time series on both regression and classification tasks.