In front of the richness of dynamical properties in neurons and central brain circuits, traditional computational architectures of artificial neuronal networks are merely based on connectivity rules. Moreover while brain circuits elaborate spike sequences, theoretical analysis usually deals with continuous signals. To understand circuit computations a different approach is needed: to elaborate realistic spiking networks and use them, together with experimental recordings of network activity, to investigate the theoretical basis of central network computation. As a benchmark we will use the cerebellar circuit. The cerebellum is supposed to work as a general purpose comparator endowed with memories and to implement forward control loops regulating movement and cognition. Experimental evidence has revealed that cerebellar circuits can dynamically regulate their activity on the millisecond scale and exploit complex spatio-temporal transformation of signals through non-linear neuronal responses and complex circuit loops. Moreover, distributed forms of plasticity can fine-tune circuit synaptic connections. In this project, we will develop specific chips and imaging techniques to perform neurophysiological recordings from multiple neurons in the cerebellar network and monitor its spatio-temporal dynamics. Based on the data, we will develop the first realistic real-time model of the cerebellum and connect it to robotic systems to evaluate circuit functioning under closed-loop conditions. The data deriving from recordings, large-scale simulations and robots will be used to explain the implicit dynamics of the circuit through the adaptable spatio-temporal filter theory. REALNET, through its network architecture based on realistic neurons, will provide a radically new view on dynamic computations in central brain circuits laying the basis for new technological applications in sensori-motor control and cognitive systems.