Title: Likelihood Based Learning Rule for Temporal Coding In Recurrent Spiking Neural Networks

Abstract: Recurrent spiking neural networks (RSNN) in the brain learn to perform a wide range of perceptual, cognitive and motor tasks very efficiently in terms of time and energy consumption. Moreover, learning can happen rapidly after very few examples. This is due to the optimality of coding and learning schemes, which have yet to be clearly understood. This naturally challenges the formulation of biologically inspired RSNNs in order to improve our understanding of biological intelligence and the efficiency of artificial ones. Several spiking network models have been proposed but it remains a challenge to design RSNNs that use biologically plausible mechanisms capable of solving complex temporal tasks. We use a general probabilistic framework that relies on the principle of maximizing the likelihood for the network to reproduce some temporal dynamics. This principle permits to analytically work out an explicit and biologically plausible plasticity rule. Here we propose a novel target-based learning scheme in which such a rule can be used to efficiently train a RSNN to solve several temporal tasks such as learning multidimensional trajectory and an implementation of the temporal XOR. We finally show that an online approximation of the gradient ascent, in addition to guaranteeing complete locality in time and space, allows learning after very few presentations of the target output. Our model is general and it can be applied to a wide variety of network architectures and types of biological neurons. The derived plasticity learning rule is specific to each neuron model and can produce a theoretical prediction to be experimentally verified.