UC San Diego

Contemporary neuroscience generates vast quantities of data through large-scale neural recordings and high-resolution behavioral tracking. Deep learning methods have emerged as a go-to for extracting understanding from increasingly complex datasets, but they resist mechanistic interpretation. At the other extreme, there is a longstanding tradition in theoretical physics of using simplified models to generate mathematically rigorous conceptual insights, yet the models often remain disconnected from empirical measurements. This dissertation explores the middle ground: parsimonious models that are both exactly solvable and can be quantitatively connected to high-dimensional, heterogeneous datasets. In the first project, we investigate the functional organization of insect primary sensory neurons. By constructing and exactly solving a minimal model of sensory neuron dynamics, we deduce that non-synaptic interactions between co-compartmentalized sensory neurons amplify the valence of olfactory stimuli, suggesting a fundamental organizing principle – the sensory periphery is structured to optimally amplify ethologically relevant odors. We validate these findings by identifying signatures of downstream connectivity patterns in the fruit-fly connectome that match model predictions. In the second project, we examine individual interactions underlying collective motion (schooling) in fish. Analysis of tracking data provides strong evidence for an analytically tractable model with stochastic, pairwise interactions. By approximately reintroducing spatial degrees of freedom in the model, we elucidate the mechanistic emergence of schooling over development, and reproduce spatial aspects of observed group dynamics. Finally, we leverage dense behavioral recordings to validate model assumptions at single swim bout resolution, and design virtual reality experiments that confirm the pairwise nature of interactions. Taken together, these projects demonstrate that confronting complex biological data does not necessarily require complex models. The success of this integrated approach – close collaboration between theory and experiments – offers a path forward for theory-driven neuroscience in an era of data abundance.