Deep nonlinear system identification models have set new standards in modeling the responses of large–scale populations of neurons to natural stimuli, yielding models that can accurately predict the response of thousands of neurons to arbitrary stimuli and can account for how behavior modulates responses of visual neuron. This allows us to treat the model as a functional digital twin of the neural population and probe neurons in ways that would not be feasible experimentally. With that, we can
derive new hypotheses about the neural populations in silico and consequently verify them in vivo, in a paradigm we call inception loops. In this talk, I will give an overview over the models, and showcase several examples how they can be used to derive novel insights for the visual system in mice and monkeys. We believe that the combination of large–scale recordings under natural stimulation and deep data–driven modeling is a paradigm shift in systems neuroscience towards understanding
computations in sensory system on complex ecological stimuli.