CENTER FOR RESEARCH IN
COGNITION & NEUROSCIENCES/ConnexConnex

Prior research on confidence in perceptual decision making has almost exclusively used simple, static stimuli and has been limited to two-choice discrimination tasks. Therefore, how confidence is given under naturalistic conditions with complex stimuli and decisions that involve multiple or even indefinite number of choices is not well understood. I will present a series of studies where we try to address this question.

Firstly, we developed a dynamic neural network model called RTNet which combines the image processing capabilities of CNNs with the empirically supported mechanism of noisy evidence accumulation. We validated this model by showing that it reproduces several important signatures of human decision making and used it to further model and compare different mechanisms of confidence generation. Model fits to human data on an eight-choice digit discrimination task showed that the best performing strategy was one which computed confidence as the difference in evidence between the top-two alternatives. These findings challenge popular views that confidence neglects all decision-incongruent information (the positive evidence heuristic) or that confidence is based on optimal computations.

In a second study, we leveraged the fact that CNNs lack built-in cognitive mechanisms to test the necessity of these mechanisms to explain human behavior. In humans, specific kinds of confidence-accuracy dissociations have typically been interpreted as evidence for confidence being influenced by additional mechanisms such as the positive evidence heuristic and noise blindness. However, CNNs, although lacking these mechanisms, were able to robustly produce these dissociations purely via changes in their perceptual representations. These findings suggest that behavioral findings cannot be automatically taken as support for cognitive processes and highlight how external, stimulus driven factors can lead to common confidence behaviors in both humans and artificial systems.

Finally, we designed a naturalistic decision-making paradigm to test how complex, real-world factors affect confidence judgments. Our findings showed that a range of factors – including affect, priors, and heuristic information – contribute selectively to confidence but not to the perceptual decision. More broadly, these results highlight major differences in how perceptual and metacognitive processes incorporate different sources of information under naturalistic conditions and begin to reveal the possible causes of inefficient metacognition observed in laboratory settings.

All together, these findings reveal the specific computations and sources of information that underlie confidence computations in natural environments and broadly support the view that metacognition arises from a more deliberative process than visual perception that integrates various sources of non-perceptual information with sensory input using heuristic computations.