Fig. 7: Alternate implementation (and generality) of the donut-like inhibitory motif.

A, B Circuit with implementation of donut-like motif purely via a reverberant route, i.e., in the absence of any feedforward inhibition. A Without recurrent amplification. B With recurrent amplification; curved orange arrow. Blue ovals: inhibitory neurons, black circles: excitatory/output neurons, dashed gray ovals: populations of neurons representing each stimulus or category. C, D Plots of CatI computed from the responses of output neuron 1 in circuits in (A) (C, filled light blue data) and in (B) (D, filled orange data), to the standard two-stimulus morphing protocol (as in Fig. 2B). For comparison, the CatI values for the corresponding models with feedforward implementation of the donut-like motif are reproduced here (in C: dark blue data from Fig. 2D; in D: red data from Fig. SH). *p < 0.05. Filled light blue vs. dark blue, p = 1.7e-25; orange vs. red, p = 2.5e-31; paired two-sided t tests with HBMC correction. Purple dashed line: CatI of purple model from Fig. 2A, which has just the feedforward donut-like motif sans feedback; reproduced here from Fig. 2D. n = 50 model neurons; center lines in the violin plots indicate median values. E–H Graphical summary of central findings of this study. E Donut-like inhibition, i.e., inhibition (−) driven by preferred inputs (+) and delivered to all non-preferred inputs, can be implemented in neural circuits either in a feedforward manner (F) or in a reverberant manner (G), to generate categorical selection boundaries (H). F, G Blue ovals: inhibitory neurons, black circles: excitatory/output neurons. F dashed thin blue arrow indicates absence of inhibitory projection. G based on Figs. 1–4; (G) is a simplified representation of model in (B), in which populations of neurons representing each category or choice (dashed gray circle) mutually inhibit one another. The similarity of this reduced model structure to several previous models of selection highlights the generality of our findings and provides a mechanistic explanation for the ability of those models to produce categorical selection boundaries^{40, 64, 67, 68}; see text. H Ability of donut-like motif to convert linear response profiles (thin black) to categorical (thick black) ones. Source data are provided as a Source Data file.