Fig. 3: Inference model theoretical predictions.

a, Simplified schema of the decision task. The orientation of the Gabor patches was drawn from the distribution of edges in natural scenes. The orientation s of the perceived Gabor patches is encoded with the internal response r. The corresponding likelihood function p(r|s) is constrained by the encoding rule. The prior f(s) is combined with the likelihood to generate an estimation \(\hat{s}\). The encoding rule depends on the model parameters q and k. b–d, Model predictions for the decision and estimation tasks assuming enhanced sensitivity k. The blue line represents a common infomax encoding model. As sensory precision increases (blue to green gradient), performance increases consistently at all orientations (b). In the estimation task, increased sensitivity k leads to a decrease in estimation bias \((\hat{s}-s)\) (c) and reduced variance over the whole range of diagonality (d). e–g, Performance predictions as a function of power-law encoding q in both tasks. As in b–d, the blue line represents the infomax model. The thick red line shows the prediction of the fitness-maximizing model. As q decreases, performance in the decision task drops for the more cardinal trials and increases for the more oblique trials (e). If q decreases for a constant capacity level, estimation biases should decrease (f). As q decreases, estimation variability decreases for oblique angles and increases for cardinal angles (g). The predictions for behaviour following an increase in sensitivity k or a decrease in the power-law encoding parameter q are thus distinctly different in terms of changes in encoding accuracy and variability.