Equivalent linear mappings of deep networks are a promising path for biology

A quantitative-genetic decomposition of a neural network

From black box to glass box\: Making UMAP interpretable with exact feature contributions