In both cortices and sensory systems, information is represented and transmitted through the correlated activity of large neuronal networks. Methods borrowed from Statistical Physics and Machine Learning are powerful tools for characterizing the collective behavior of large systems and thus offer promising approaches to understand the activity of neuronal populations. In this talk I will show how the Maximum Entropy principle, applied to cortical in-vivo recording, allows for comparing the population behavior during wakefulness and deep sleep and eventually for identifying cell-assemblies, i.e. strongly co-activated groups of neurons that play a central role in memory consolidation. I will then use hidden layer models, point processes and “experimental” linear response theory to account for the non-linear stimulus processing in sensory networks such as the retina. These approaches allow for constructing high performing models of the retinal population response to visual stimuli and thus for characterizing how a network of neurons can encode and transmit visual information.