Figure 38. Effects of segregated or mixed-modality signals.

(A) Schematic of two possible architectures for an MB-like circuit, in which KCs are either specialized for one sensory modality (left) or receive mixed input (right). Bottom layer circles represent PN inputs, middle layer circles represent KCs, and black output circles represent MBONs. Green and yellow shades indicate different sensory modalities. (B) Task performance in a stimulus discrimination model for the architectures shown in A. Stimuli consist of random binary patterns in two sensory modalities. In the ‘factorizable’ task, the stimulus from each modality has an assigned valence (positive or negative) and the overall valence is positive only if both constituent valences are positive. In the ‘unfactorizable’ task, each stimulus is randomly assigned its own unique valence, irrespective of its constituent modalities. A linear readout, modeling an MBON, is fit to discriminate the valences of the stimuli based on the activations of the KC population. The linear fit procedure corresponds to the experienced-based plasticity of KC-MBON synapses. Task performance is defined as the binary classification accuracy—that is, the frequency with which the thresholded response of the MBON corresponds correctly to the valence of the input—of this MBON readout. 25 stimuli are presented in the unfactorizable task and 200 in the factorizable task, to make overall task difficulty comparable. Error bars represent standard error over 20 simulations. (C) Schematic of possible architectures for DAN-dependent plasticity in an MB-like circuit. In the leftmost model, DANs (colored squares) modulate KC-MBON synapses for KCs of a particular sensory modality. The second-to-left model depicts two MBONs, each integrating input from KCs of a particular sensory modality, and each with its own corresponding DAN, that additively converge onto a common output (blue). These two models are functionally equivalent, and hence are grouped together under ‘separate reward signals.’ In the rightmost model, there is a single MBON, and DANs modulate all the KC-MBON synapses (right). (D) Task performance in a stimulus discrimination model (same as B) for the architectures shown in C. For each task the optimal KC structure from B is used (‘separate’ for the factorizable task, ‘shared’ for the unfactorizable task).