Fig. 1: Precision vs. global pooling and credit assignment.

a Precision pooling: in precision pooling, only a subset of neurons of a population is read out. For conjunction learning, one option is to read out only neurons purely tuned to color and to orientation, respectively. Their output can be summed to give rise to conjunction information (top left panel). Alternatively, only neurons with mixed selectivity for color and orientation can be read out. These neurons provide information about conjunctions directly and explicitly (top-right panel). Global pooling: In global pooling, all neurons are read out regardless of their tuning properties. Hence, neurons purely tuned to color and to orientation, as well as neurons with mixed selectivity for both features, contribute to conjunction learning. b Credit assignment regimes in conjunction learning with global or feature-specific feedback. When feedback is global (on the conjunction level), the three relevant resources, pure orientation, pure color, and mixed selectivity, are up-/down-weighted together. However, when feature-specific feedback is provided, credit is only consistently assigned to pure selectivity neurons: when one feature is correct and the other incorrect, credit cannot be unambiguously assigned to neurons with mixed selectivity (signified by black lines on up and down-weighting arrows in the middle column). Hence, their weights are updated less frequently (or consistently) than those of pure selectivity neurons, i.e., only when the response is correct or incorrect on both features. Green and red colors in the “feedback” column represent correct and error feedback, respectively. Green upward and red downward arrows indicate up/down weighting of relevant resources following correct and error feedback, respectively.