Analysis of Learning and Inference in Spiking Neurons via Synaptic Plasticity

Solving classification problems with spiking neural networks offers significant advantages in terms of energy efficiency and computational cost. However, training a spiking neural network to solve a specific problem is quite challenging. Despite the development of various biology-based approaches, learning is not as straightforward and mathematically traceable as in artificial neural networks. In such networks, neurons are typically trained unsupervised with spike-time-dependent synaptic plasticity. To have control over the overall behavior of the network, it is important to master the dynamics of a single neuron. In this study, the learning and inference stages were examined using a single spiking neuron model and a single dataset. The study has shown that spiking neurons demonstrated high success in learning the patterns presented as input. Simultaneously, we concluded that these neurons proved inadequate during the inference stage, where we attempted to make predictions. Also, the analysis of the input data revealed that the neuron models that performed well and the training based on synaptic plasticity are greatly affected by biases in the data.