Progress in understanding brain functions rely in great part on filling the conceptual and experimental gaps between different levels of analysis, from single neurons to behaviour. Thus, “rate models”, units of representations are the mean activity of large neural populations, while function and behaviour emerge from the responses of very large networks. While experimental investigations have focused on predicting (describing) spiking neural responses from their (sensory or synaptic) inputs, functional models instead concentrate on understanding how neural populations represent properties of (i.e. predict) the world.This proposal aims at developing an alternative approach, spike-based predictive coding. It combines two basic hypotheses: Neural networks reliably estimate the state of the environment based on their inputs and prior experience. And their dynamics insures that these estimates can be decoded from their spike trains by postsynaptic integration . By monitoring and decoding its own outputs, the neural structure itself closes the loop between computation and dynamics.Membrane potentials of model neurons compute a difference between the state estimates constructed from their inputs and the estimate encoded in their outputs. Interestingly, this purely functional approach converges towards powerful descriptive models of spiking neurons, e.g. adaptive integrate and fire neurons, chaotic attractors in balanced spiking networks and generalized linear models (GLMs).We will use this approach to explore the dynamics of single spiking neurons, suggest new ways of interpreting and exploring sensory and motor spiking neural representations, re-explore the role of top-down attention in sensory processing, and show that previous rate-based interpretations severely under-estimated the precision of the neural code.