An Iterative Framework for Assessing the Firing Probability of Leaky Integrate and Fire Neurons With a Novel Threshold Adaptation Mechanism

Artificial intelligence and deep learning have opened a new frontier for solving complex real-life problems that are not analytically solvable. The core neuron model in the deep neural network (DNN) is based on activation functions. Recently, a shift has been observed towards biologically plausible, energy-efficient spiking neural networks (SNNs) that use spiking neurons as their core computational units. Leaky integrate-and-fire (LIF) models are widely used in spiking neural networks. However, a simple LIF model lacks spike-frequency adaptation (SFA), an intrinsic property of biological neurons. SFA provides spiking stability and prevents excessive spike generation under sustained stimulation. One way to implement SFA in spiking neuron models is through a history-dependent adaptive threshold. A novel neuron model based on the LIF framework with a prior spiking history-based threshold adaptation technique is proposed in this paper. Unlike conventional adaptive LIF models that use a fixed threshold increment after each spiking event, the proposed model employs inter-spike-interval-based threshold adaptation. The model is compared with three closely related existing models through extensive simulations, which show that the proposed model attains a better SFA. The proposed Spiking neuron model with the novel threshold adaptation technique has been used to perform an iterative statistical analysis of the probability of spike generation over time. A non-stationary random process has been treated as a sequence of random variables through this iterative statistical framework.