Coral reefs are among the most biodiverse ecosystems on Earth in large part owing to their unique three-dimensional (3D) structure, which provides niches for a variety of species. Metrics of structural complexity have been shown to correlate with the abundance and diversity of fish and other marine organisms, but they are imperfect representations of a surface that can oversimplify key structural elements and bias discoveries. Moreover, they require researchers to make relatively uninformed guesses about the features and spatial scales relevant to species of interest. This paper introduces a machine-learning method for automating inferences about fish abundance from reef 3D models. It demonstrates the capacity of a convolutional neural network (ConvNet) to learn ecological patterns that are extremely subtle, if not invisible, to the human eye. It is the first time in the literature that no a priori assumptions are made about the bathymetry–fish relationship.