Interneuron diversity and normalization specificity in a visual system

Abstract

AbstractNormalization is a fundamental operation in image processing. Convolutional nets have evolved to include a large number of normalizations (Ioffe and Szegedy 2015; Ulyanov, Vedaldi, and Lempitsky 2016; Wu and He 2018), and this architectural shift has proved essential for robust computer vision (He et al. 2015; Bjorck et al. 2018; Santurkar, Tsipras, and Ilyas 2018). Studies of biological vision, in contrast, have invoked just one or a few normalizations to model psychophysical (Mach 1868; Furman 1965; Sperling 1970) and physiological (Carandini and Heeger 2011; Shin and Adesnik 2024) observations that have accumulated for over a century. Here connectomic information (Matsliah et al. 2023) is used to argue that interneurons of the fly visual system support a large number of normalizations with unprecedented specificity. Ten interneuron types in the distal medulla (Dm) of the fly optic lobe, for example, appear to support chiefly spatial normalizations, each of which is specific to a single cell type and length scale. Another Dm type supports normalization over features as well as space. Two outlier types do not appear to support normalization at all. Interneuron types likely to be normalizers are identified not only in Dm but also in all other interneuron families of the optic lobe. For fly vision, the diversity of interneurons appears to be an inevitable consequence of the specificity of normalizations.