The study introduces a novel modeling approach, "knockout training," to identify a one-to-one mapping between internal units in a deep neural network (DNN) and real neurons in the brain of *Drosophila melanogaster*. This approach involves perturbing the network during training to match the perturbations of real neurons during behavioral experiments. The researchers applied this method to model the sensorimotor transformations of male flies during complex, visually guided social behaviors, such as courtship. They found that combinations of visual projection neurons, including those involved in non-social behaviors, drive male interactions with females, forming a rich population code for behavior. The model not only predicts how individual neurons contribute to behavior but also infers neural activity from perturbed behavior alone, making it useful for studying complex, natural behaviors. The study provides a framework that consolidates behavioral effects elicited from various neural perturbations into a single, unified model, enabling future incorporation of wiring diagrams of the brain into the model.The study introduces a novel modeling approach, "knockout training," to identify a one-to-one mapping between internal units in a deep neural network (DNN) and real neurons in the brain of *Drosophila melanogaster*. This approach involves perturbing the network during training to match the perturbations of real neurons during behavioral experiments. The researchers applied this method to model the sensorimotor transformations of male flies during complex, visually guided social behaviors, such as courtship. They found that combinations of visual projection neurons, including those involved in non-social behaviors, drive male interactions with females, forming a rich population code for behavior. The model not only predicts how individual neurons contribute to behavior but also infers neural activity from perturbed behavior alone, making it useful for studying complex, natural behaviors. The study provides a framework that consolidates behavioral effects elicited from various neural perturbations into a single, unified model, enabling future incorporation of wiring diagrams of the brain into the model.