Figure 1. Distractor effects predicted by mutual inhibition model, divisive normalisation model, and a dual route model that combines both other models.

(a) A mutual inhibition model involves three pools of excitatory neurons that receive excitatory input from the HV, LV or D options (P_HV, P_LV and P_D). Concurrently, all excitatory neurons further excite a common pool of inhibitory neurons P_Inh, which in turn inhibit all excitatory neurons to the same extent. The HV or LV option is chosen once its accumulated evidence (y_HV or y_LV respectively) reaches a decision threshold. (b) The decision accuracy of the model is plotted across a decision space defined by the difficulty level (i.e. value difference between HV and LV) and the relative distractor value (D–HV). The model predicts a positive distractor effect – the decision accuracy increases (brighter colors) as a function of relative distractor value (left-to-right side). (c) A positive distractor effect is found on both hard (bottom) and easy (top) trials. (d) A GLM analysis shows that the model exhibits a positive HV-LV effect, a positive D-HV effect and a positive (HV-LV)(D–HV) effect. (e) Alternatively, a divisive normalisation model involves only two pools of excitatory neurons that receive input from either the HV or LV option. The input of each option is normalised by the value sum of all options (i.e. HV+LV+D), such that the distractor influences the model’s evidence accumulation at the input level. (f) Unlike the mutual inhibition model, the divisive normalisation model predicts that larger distractor values (left-to-right side) will have a negative effect (darker colours) on decision accuracy. (g) A negative distractor effect is found on both hard (bottom) and easy (top) trials. (h) A GLM analysis shows that the model exhibits a positive HV-LV effect, a negative D-HV effect, and a negative (HV-LV)(D–HV) effect. (i) A dual route model involves evidence accumulation via mutual inhibition and divisive normalisation components independently. A choice is made by the model when one of the components accumulates evidence that reaches the decision threshold. (j) The current model predicts that on hard trials (bottom) larger distractor values (left-to-right side) will have a positive effect (brighter colors) on decision accuracy. In contrast, on easy trials (top) larger distractor values will have a negative effect (the colors change from white to yellow from left to right). (k) The opposite distractor effects are particularly obvious when focusing on the hardest (bottom) and easiest (top) trials. (l) A GLM analysis shows that the model exhibits a positive HV-LV effect, a positive D-HV effect and a negative (HV-LV)(D–HV) effect.

Figure 1—figure supplement 1. In the dual route model the positive D-HV effect on hard trials is mainly contributed by the mutual inhibition component and the negative D-HV effect on easy trials is mainly contributed by the divisive normalisation component.

(a,b) Replica of Figure 1j and k showing choice accuracy predicted by the dual route model as a function of relative distractor value D-HV and difficulty level. (c) The choices made by the mutual inhibition (Mutual) component and divisive normalisation (DivNorm) components are analysed separately. On hard trials (bottom), accuracy of choices made by the Mutual component (blue lines) increases as a function of D-HV. A similar trend is observed on easy trials (top), but with a smaller slope. In contrast, accuracy of choices made the DivNorm component (red lines) decreases as a function of D-HV in both hard and easy trials and the negative slope is steeper on easy trials. (d) There are overall four possible types of choices made by the dual route model – accurate and error choices made by the Mutual component and accurate and error choices made by the DivNorm component. (Bottom) On hard trials, in choices made by the Mutual component the proportion of errors (light blue bars) decreases much more rapidly than the accurate choices (dark blue bars) as a function of D-HV. In choices made by the DivNorm component, the proportions of accurate (red) and error (pink) choices made by the divisive normalisation model increase to a similar extent. Thus, there is an overall net increase in accuracy on hard trials as a function of D-HV. (Top) In contrast, on easy trials errors made by the Mutual component are rare and there is little variability as a function of D-HV. However, errors made by the divisive normalisation component increase as a function of D-HV. Thus, there is an overall net decrease in accuracy on easy trials as a function D-HV. The green lines show the proportion of overall accurate choices, which are identical to the lines on panel b.

Figure 1—figure supplement 2. Reaction time (RT) of choices made by the dual route model.

(a) RT of the mutual inhibition component when the divisive normalisation component is switched off. (b) RT of the divisive normalisation component when the mutual inhibition component is switched off. Although the RT of the mutual inhibition component is generally faster than that of the divisive normalisation component, due to variability in the RTs of both components, choices are sometimes made by the divisive normalisation component (see Figure 1—figure supplement 1d).