A study published in Nature (https://doi.org/10.1038/s41586-024-07451-8) presents a novel approach to map internal units in a deep neural network (DNN) to real neurons in the brain of Drosophila melanogaster males. The researchers developed a method called 'knockout training' to identify a one-to-one mapping between DNN units and real neurons by simulating the effects of silencing different neuronal cell types. This approach was applied to model the sensorimotor transformations during a complex, visually guided social behavior in male flies, specifically courtship behavior. The study reveals 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 DNN model, called the '1-to-1 network', was trained to predict the behavioral changes resulting from systematic perturbations of more than a dozen neuronal cell types. The model successfully captures the behavioral repertoire of genetically altered flies when the corresponding model units are silenced, aligning the model units with real neurons. The 1-to-1 network was tested against real LC neuron responses and showed that its predicted responses largely matched the corresponding real LC responses for artificial stimuli. The model also accurately predicted responses to naturalistic stimulus sequences, demonstrating its ability to decode visual features and behavioral outputs. The study further shows that the model LC units encode visual stimuli in a distributed way, with each visual stimulus feature encoded by multiple model LC units and each model LC unit encoding multiple visual stimulus features. The research highlights the importance of population coding in behavior, particularly in natural contexts, and suggests that the complex courtship behavior of the male relies on combinations of visual projection neurons, including those also involved in non-social behaviors. The study also discusses the implications of the findings for understanding the neural basis of behavior and the potential for future experiments to test the predictions of the 1-to-1 network.A study published in Nature (https://doi.org/10.1038/s41586-024-07451-8) presents a novel approach to map internal units in a deep neural network (DNN) to real neurons in the brain of Drosophila melanogaster males. The researchers developed a method called 'knockout training' to identify a one-to-one mapping between DNN units and real neurons by simulating the effects of silencing different neuronal cell types. This approach was applied to model the sensorimotor transformations during a complex, visually guided social behavior in male flies, specifically courtship behavior. The study reveals 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 DNN model, called the '1-to-1 network', was trained to predict the behavioral changes resulting from systematic perturbations of more than a dozen neuronal cell types. The model successfully captures the behavioral repertoire of genetically altered flies when the corresponding model units are silenced, aligning the model units with real neurons. The 1-to-1 network was tested against real LC neuron responses and showed that its predicted responses largely matched the corresponding real LC responses for artificial stimuli. The model also accurately predicted responses to naturalistic stimulus sequences, demonstrating its ability to decode visual features and behavioral outputs. The study further shows that the model LC units encode visual stimuli in a distributed way, with each visual stimulus feature encoded by multiple model LC units and each model LC unit encoding multiple visual stimulus features. The research highlights the importance of population coding in behavior, particularly in natural contexts, and suggests that the complex courtship behavior of the male relies on combinations of visual projection neurons, including those also involved in non-social behaviors. The study also discusses the implications of the findings for understanding the neural basis of behavior and the potential for future experiments to test the predictions of the 1-to-1 network.