Welcome to Deep Brain Discovery a.k.a. DeepBrainDisco!
Models of neural architecture and organization are critical for the study of disease, aging, and development. In our work Deep Brain Discovery, we leverage deep neural networks to build robust and expressive data-driven models of brain structure across multiple, diversified brain regions. In particular, we show how low-dimensional matrix factorization can be applied to network activations to learn new representations of the data that possess more interpretable structure. We show that our newly learnt representations have components that are more biologically meaningful and therefore better suited for modeling neural architecture. Through multiple vignettes, we present ways to utilize these features for the discovery of anatomically linked macrostructure, new regions of interest, and anatomical motifs within the sample.