Interneuron diversity and normalization specificity in a visual system
Normalization is a fundamental operation in image processing. Convolutional nets have evolved over the past decade to include a large number of normalizations, and this architectural shift has proved essential for robust visual artificial intelligence. Here I argue that normalization is the function of a large fraction of neuronal cell types in the fly optic lobe. A dozen types of local interneuron in the distal medulla (Dm), in particular, are argued to mediate lateral inhibition, which was classically assumed to depend only on spatial separation. Each Dm type turns out to connect with specific source and target cell types. The source and target are either the same type, or are directly connected, from which it follows that Dm-mediated lateral inhibition has a normalizing function. Therefore the diversity of Dm interneurons is analogous to the ubiquity of normalizations in contemporary convolutional nets. A notable difference is that Dm types normalize over specific spatial scales. A final Dm type is an outlier in its perfect tiling of the medulla, and likely does not mediate lateral inhibition due to a predicted electrical compartmentalization of its arbor. Further candidate normalizers are identified in all other interneuron families, generalizing the Dm findings to the entire optic lobe.