Characterizing feature selectivities of sensory visual neurons is crucial for understanding how the brain forms representations of the external world. In vivo neurophysiological experiments aimed at identifying such tuning properties are constrained by time, making exhaustive searches in stimulus space infeasible. In recent years, deep neural networks (DNNs) have set new standards in predicting neuronal responses of the visual system to arbitrary stimuli, making in silico exploration of the stimulus-response space possible. As visual input is dynamic, there is a need for models that can faithfully predict responses to both natural movies and parameterized synthetic stimuli, as used in many classical neurophysiological experiments. The generalization from natural movies to parameterized synthetic stimuli is a precondition for DNNs to become useful at systematic in silico studies of feature selectivities. Here, we investigated this precondition for a data-driven dynamic DNN model of retinal ganglion cell (RGC) axonal boutons in the superior colliculus (SC) of mice (Fig.1A). Our model captured both visual stimulus specific components of the neuronal responses (correlation to average across repeated trials) and trial-to-trial variabilities (single-trial correlation) with high accuracy. We also find that our DNN generalizes from natural movies to parameterized synthetic stimuli commonly used during neurophysiological experiments, such as chirp and moving bar (Fig.1B,C). Analyzing the predicted responses to the latter we successfully identified direction and orientation selective RGC axonal boutons (Fig.1D). We further investigated the functional organization of direction and orientation tuned RGC axonal boutons in the SC as predicted by the model. We observed that organization of direction and orientation selective SC neurons into patches and columns, respectively [1,2], was present already at the level of RGC axonal boutons (Fig.1E,F), as reported in recent experimental findings on the functional architecture of RGC axonal boutons in the SC [3]. This demonstrates the capability of our dynamic DNN not only for in silico characterizing neuronal tuning properties, but also for accurately capturing their spatial relations into functional maps.