Abstract : Recent experimental advances have made it possible to record several hundred neurons simultaneously in the retina as well as in the cortex. Analyzing such a huge amount of data requires to elaborate statistical, mathematical and numer- ical methods, to describe both the spatio-temporal structure of the population activity and its relevance to sensory coding. Among these methods, the maxi- mum entropy principle has been used to describe the statistics of spike trains. Recall that the maximum entropy principle consists of xing a set of constraints, determined with the empirical average of quantities ("observables") measured on the raster: for example average ring rate of neurons, or pairwise corre- lations. Maximising the statistical entropy given those constraints provides a probability distribution, called a Gibbs distribution, that provides a statistical model to t the data and extrapolate phenomenological laws. Most approaches were restricted to instantaneous observables i.e. quantities corresponding to spikes occurring at the same time (singlets, pairs, triplets and so on).