Valence, Not Utility, Underlies Reward-Driven Prioritization in Human Vision

Abstract

Objects associated with reward draw attention and evoke enhanced activity in visual cortex. What is the underlying mechanism? One possibility is that reward's impact on vision is mediated by unique circuitry that modulates sensory processing, selectively increasing the salience of reward-associated stimuli. Alternatively, effects of reward may be part of a more general mechanism that prioritizes the processing of any beneficial object, importantly including stimuli that are associated with the evasion of loss. Here, we test these competing hypotheses by having male and female humans detect naturalistic objects associated with monetary reward, the evasion of equivalent loss, or neither of these. If vision is economically normative, processing of objects associated with reward and evasion of loss should be prioritized relative to neutral stimuli. Results from fMRI and behavioral experiments show that this is not the case: whereas objects associated with reward were better detected and represented in ventral visual cortex, detection and representation of stimuli associated with the evasion of loss were degraded. Representations in parietal cortex reveal a notable exception to this pattern, showing enhanced encoding of both reward- and loss-associated stimuli. Experience-driven visual prioritization can thus be economically irrational, driven by valence rather than objective utility.SIGNIFICANCE STATEMENT Normative economic models propose that gain should have the same value as evasion of equivalent loss. Is human vision rational in this way? Objects associated with reward draw attention and are well represented in visual cortex. This is thought to have evolutionary origins, highlighting objects likely to provide benefit in the future. But benefit can be conferred not only through gain, but also through evasion of loss. Here we demonstrate that the visual system prioritizes real-world objects presented in images of natural scenes only when these objects have been associated with reward, not when they have provided the opportunity to evade financial loss. Visual selection is thus non-normative and economically irrational, driven by valence rather than objective utility.

Keywords: MVPA; attention; fMRI; incentive salience; reward; vision.

Figure 1.

Predictions from the utility and valence models. A, By the utility model, reward- and loss-associated targets should draw selective resources, and thus be better represented than neutral targets. B, During search for neutral targets, reward- and loss-associated distractors should require the same degree of attentional suppression, and thus be poorly represented relative to neutral distractors. C, By the valence model, reward-associated targets should draw selective resources, whereas loss-associated targets may be actively suppressed and poorly represented in the visual system. D, Reward-associated distractors should require attentional suppression, but loss-associated distractors should not.

Figure 2.

A, Schematic illustration of the trial sequence. Participants reported the presence of examples of the cued category in briefly presented scenes. Of four possible target categories: one was associated with reward, one to loss, and two to neutral outcome. Image and font sizes are not to scale, and the block cue here indicates only the target category for the coming trials, whereas in the experiment itself the total number of points earned to that point in the experiment was also presented. B, Feedback schedule. The association of category to outcome in the actual experiment was counterbalanced across participants. Feedback indicated here was for target-present trials. Correct performance in target-absent trials garnered a single point in neutral blocks and 0 points in reward and loss blocks. Incorrect performance in target-absent trials resulted in the loss of a single point in neutral blocks and the loss of 50 points in reward and loss blocks. C, Analytic approach. Scene-evoked activity patterns in OSC were correlated with benchmark patterns. High correlation indicates increased information for that category in visual cortex during scene perception.

Figure 3.

Results from Experiment 1. A, In line with the valence model, reward-associated targets are better represented in OSC than loss-associated targets. B, During search for neutral targets, OSC carried less information about reward-associated distractors than loss-associated distractors, indicative of attentional suppression. Our normalization procedure causes these values to be represented on an interval scale with an uninformative zero point. As such, negative values do not imply the presence of information in the form of negative correlation (for details, see Materials and Methods). C, Accuracy in detecting the target when it was present in the scene. D, The relationship between coefficients from a fit of the valence model to OSC category information, as illustrated in A, B, and coefficients from a fit of the valence model to hit rate data, as illustrated in C. Error bars indicate within-participant SE (Cousineau, 2005).

Figure 4.

A, Anatomically defined ROIs characterizing the SN and red nucleus. B, Relationship between interindividual variance in the valence coefficient of the SN ROI and the valence coefficient of category information in OSC. C, Midbrain voxels identified in whole-brain correlational analysis. The single 3 mm voxel on the midline actually begins one slice inferior to the illustrated horizontal section (beginning at z = −14). D–F, Other clusters identified in whole-brain correlational analysis. G, Relationship between interindividual variance in BIS and the valence coefficient of category information in OSC.

Figure 5.

A, OSC as defined in the OSC localizer. Voxels identified here were present in the OSC of 16 or more of the 23 participants in Experiment 1. B, Results from the searchlight contrast of information for targets versus distractors. Voxels identified here constitute the center of spheres that carried more information for targets than distractors at p < 0.001 with a cluster threshold of 50 voxels.

Figure 6.

A, Parietal cluster identified in searchlight analysis. This brain region was defined by contrasting information content for targets versus distractors. Centroid: 9, −73, 43, MNI space. B, Results from analysis of outcome association in the parietal cluster. Error bars indicate within-participant SE (Cousineau, 2005).

Figure 7.

Results from Experiment 2. A, In line with the valence model, accuracy is better for reward-associated targets than loss-associated targets. B, Search for a neutral target is more strongly disrupted by a reward- versus loss-associated distractor.

Figure 8.

Further analysis of results from Experiment 2. A, Explicit accuracy in target-present trials as observed when a response was made within the time limit. B, Explicit accuracy in target-absent trials. C, Perceptual sensitivity for targets. D, Response criterion. Participants tended to report the target as present, resulting in a negative criterion, but this did not differ across reward and loss conditions. E, Reaction times for target-present and target-absent trials as a function of target association. F, Reaction times for target-present and target-absent trials where the target had neutral association, as a function of whether a reward-, loss-, or neutral-associated distractor was present in the scene. G, Number of time-out trials when target was present, as proportion of total per condition. H, Number of time-out trials when target was absent. Error bars indicate within-participant SE (Cousineau, 2005).