Membrane-potential-dependent plasticity learning for theta-neuron network

In this work, we propose and test a procedure to guide learning in spiking neural networks using two layers of inhibitory and excitatory Ermentrout-Kopell canonical model (theta) neurons. A wide range of neuronal dynamics, including low-amplitude oscillations, is known to influence information processing. Current implementations of Hebb theory-based algorithms, such as the spike-timing-dependent plasticity (STDP), fail to account for these, and most formulations of the STDP being discrete are not biologically feasible either. Our membrane potential-dependent plasticity (MPDP) algorithm necessarily includes both spiking and sub-threshold neuronal activity, and we use a continuous formulation of the MPDP to perform unsupervised pattern recognition in a spiking neural network. To the best of our knowledge, this is the first Hebb theory-based learning algorithm for theta neurons. The general formulation that we propose renders it accessible for other complex neuron models, such as the FitzHugh-Nagumo and the Hindmarsh-Rose models, as well as systems of coupled neurons. Our results suggest that the average membrane potential captures features salient for information decoding that are overlooked when considering only the pure spiking rate.