Figure 2.

Functional architecture. The input field (F1) consists of binary feature detectors: a particular input consisting of five active features is shown. The coding field, F2, is proposed as a macrocolumn analog. It consists of winner-take-all competitive modules (CMs) proposed as analogous to minicolumns. Each CM has K = 3 binary units. Bottom-up connectivity is all-to-all: gray lines signify initially 0 weights. The familiarity (G) of the input is proposed to be computed via a subcortical, neuromodulator-based mechanism, which then modulates the F2 unit activation function parameters (e.g., sigmoid height) contingent on G (purple arrows). See text.