A novel classification technique using a biologically plausible spiking neuron and noisy synapses

The organic evidence from neuroscience proves that precise spike times are used for information exchange between two biological neurons rather than the firing rates. One of the prominent reasons, along with energy and computational efficiency, is that spiking neural networks (SNNs) are getting more attention nowadays. The spiking neurons in SNN mimic the biological neuron more than its predecessors. Despite the few efficient supervised learning algorithms for SNN, only some investigated the biological properties such as axonal noise, random synaptic delays, spontaneous spike-firing, and random switching of the gamma-aminobutyric acid (GABA)-switch. The aforementioned properties are essential for making spiking neurons more biologically realistic, which is one of the major strengths of SNN. The GABA switch decides the most crucial activity, whether a neuron will be excitatory or inhibitory. This paper proposes a novel and efficient approach to handle non-linear patterns using a single leaky-integrate-and-fire (LIF) spiking neuron connected with many noisy synapses with random synaptic delays. In addition, the spontaneous firing of a neuron and random switching of signs in synaptic weights having equal probability akin to GABA-switch are efficiently implemented. Moreover, a hybrid kernel is proposed as the synapse model to cope with the noise properly, which makes the synapse model more efficient. The error-tuning is carried out using the elitist floating-point genetic algorithm. Four datasets were used for benchmarking, and experimentally, better results were obtained than state-of-the-art methods.