An important problem in systems neuroscience is to understand how animals transform complex, high-dimensional sensory inputs into neural signals that drive behavior. Here we propose a novel modeling approach for this sensorimotor transformation to identify a one-to-one mapping between real and model neurons. Our approach involves training a deep neural network to produce behavioral data from a range of different inactivation experiments. Crucially, we train the model while perturbing its hidden units with the same perturbations applied to neural activity—a process we call ‘knockout’ training. We applied this approach to understand the sensorimotor transformation of Drosophila melanogaster males during a complex, visually-guided social behavior. We explicitly modeled a population of visual projection neurons known as Lobula Columnar or LC cells (Wu et al., eLife 2016), which receive inputs from the optic lobes and send their axons to the central brain. Our model identified a one-to-one mapping from model neurons to LC cells, providing an overview into how LCs coordinate their activity to transform visual input to drive behavior. We verified that the model correctly predicted responses recorded from LC neurons—even though the model had no access to neural recordings during training. We then systematically analyzed our model and found that most of the model LCs had mixed selectivity and contributed to multiple behavioral actions (e.g., movement and song production). This suggests that, contrary to the current prevailing view, LC neurons form a distributed population code to sculpt social behavior. Overall, we propose a novel modeling framework for relating deep neural network models to real neurons and to shed light on the neural computations performed during sensorimotor transformations.