New methods to identify one-to-one mappings between neural network models and brain neurons

About This Grant

To understand how a neuron contributes to behavior, neuroscientists rely on a diverse set of tools and techniques to perturb neural responses and observe changes in behavior. As the amount of neural perturbation data rapidly grows, new methods are needed to build models that both incorporate and explain this neural perturbation data. We recently developed knockout training, a new machine learning method that, at its core, perturbs or “knocks out” model units of neural network models in the same way neurons are inactivated in the brain; models are trained on behavioral data from animals with neural perturbations, and the models learn how neurons both represent sensory information and drive motor behaviors. We will expand and improve this method, and thereby generate general-purpose knockout training algorithms widely applicable and available to the systems neuroscience community. To evaluate knockout training’s ability to estimate one-to-one mappings, we model neural systems in Drosophila melanogaster, chosen for its highly-structured anatomy, wealth of genetic and perturbation tools, rich natural behaviors, and available connectomes. An important advance of the proposed research is to apply knockout training to biologically-realistic models, leveraging detailed anatomical data from connectomes. The specific aims include: i) improving knockout training algorithms for settings with little training data, access to neural recordings, and large behavioral stochasticity, and evaluating these algorithms with extensive simulations and real perturbation data targeting the Drosophila visual system; ii) leveraging recently released connectomic data to build connectome-informed models of the Drosophila visual system, which we then train with extensions of knockout training tailored to biologically-realistic models; And, iii) evaluating the proposed knockout training algorithms on a connectome-informed model that includes both the Drosophila visual and premotor systems. The end result of this work will be a suite of well-tested, general-purpose knockout training methods able to integrate perturbation data with deep neural network and biologically-realistic models of the brain, ready for wide adoption by the systems neuroscience community. This work will also substantially advance our understanding of the neural circuits and computations involved in sensorimotor transformations. Our knockout-trained, connectome-informed model of the fruit fly visual and motor systems will act as a large-scale, working hypothesis of the fruit fly brain that the Drosophila community may use to guide future experiments. Our framework consolidates the effects of many individual perturbations into one unified model; we can then perform experiments on the model—silencing combinations of neuron types—not possible to perform on real animals. We foresee methods that incorporate perturbation data into computational models, such as the knockout training algorithms proposed here, becoming as invaluable as the perturbation techniques themselves. This is especially true when studying larger animals with more complex behaviors and with neural systems that may have tens of thousands of neuron types to perturb.