Article Preview
Top1. Introduction
Human brain structure and its working is very complicated. Hence, simulation of the working of human brain is a challenging research area. Brain receives different sensory inputs from the environment, processes these inputs itself, and performs various actions accordingly. It looks so simple in our daily routine, but a very complicated neural structure is responsible for this. Millions of neurons connect with each other to process the inputs, which may vary with time and decide an action accordingly. The responses of the neurons to the same external stimuli may also vary. Biological neurons communicate with each other by generating electrical pulses. These pulses are known as spikes or action potential that travel from one neuron to another (Vreeken, 2003). Whenever, an action potential creates a change in the membrane potential of a target neuron and this change crosses a threshold, a new spike is generated.
Spiking Neural Networks (SNN) are the closest simulation of brain functioning since the communication between the spiking neurons is in the form of spikes (Bohte, 2004; Gerstner et al., 2002). Spikes are an essential aspect of SNN. The input spike train are fed to a SNN and the output neuron emits a spike if the output threshold is crossed. Due to the computational and biological characteristics of the SNN, they are used for various engineering applications. Temporal information is stored in the form of number of spikes generated and the precise timing of these spikes(Cessac et al., 2010). In addition, the firing response of the spiking neuron also varies in response to input signal fluctuations over time. The efficiency of a SNN is measured quantitatively by counting the number of spikes and qualitatively by the variety of firing responses of the spiking neuron to time-varying input currents (Paugam-Moisy & Bohte, 2012; Yamauchi et al., 2011). Multi-timescale adaptive threshold (MAT) model was deisgned by Yamauchi et al., which can generate a variety of firing responses (Yamauchi et al., 2011). Ori et al. present the dynamic clamp constructed phase diagram that represents the relationship between sodium and potassium channel parameters and neuron membrane excitability using HH model (Ori et al.2020). Alonso and Marder developed a new method by using simple mathematical concepts to display ionic currents in computational models of neurons (Alonso & Marder, 2019). Campbell used the synthesis dataset for HH model experiments (Campbell, 2020).
In the general process of spiking, action potentials travel with axons and activate the synapses (Colwell & Brenner, 2009). These synapses release a neurotransmitter that quickly diffuses to the postsynaptic neuron. In the postsynaptic neuron, these neurotransmitters affect the neuron’s membrane potential. Excitatory Postsynaptic Potentials (EPSPs) increase the membrane potential (depolarize), and without new inputs, this excitation then leaks away with a typical time constant. Inhibitory Postsynaptic Potentials (IPSPs) slow down the membrane potential (hyperpolarization). When enough EPSPs arrive at a neuron, the membrane potential depolarizes enough to reach a certain threshold. At that time, the neuron itself generates a spike to reset its membrane potential. Due to this process, the generated spike then goes to other neurons. Recent research trends show that neurons encode information in the timing of single spikes, and not only just in their average firing frequency. Several researchers have proved that spiking neurons are more powerful than their non-spiking counterparts such as feed-forward back-propagation multilayer perceptron neural network (Susan et al., 2019) because they can also encode temporal information with their signals (Grüning & Bohte, 2014; Ponulak & Andrzej, 2011). Time-varying signals are processed and classified in convention through the formulation of temporal features (Susan et al. 2019). In SNN, the time information is incorporated into the functionality of the basic unit itself which is the spiking neuron. Hence SNN is the form of neural network most suited for analysis of time-series signals