Adam Kepecs, PhD
Professor, Neuroscience Chair
Department of Neuroscience
Cold Spring Harbor Laboratory
(October 29, 2019)
Neurobiology of confidence: statistics, neurons, and psychiatry
Consider a time when you felt very confident in a decision or in an answer on a test. Before you knew the outcome, you felt almost certain that you were correct. How did that certainty affect your next decisions or your next test answers? Dr. Kepecs discussed work in humans and animals showing that confidence plays a role in discriminability task performance. Neurons that respond in reference to confidence are found in the lateral orbitalfrontal cortex of rats, as well as the ventral tegmental area, and are important for reinforcement learning in rats.
Confidence is a term with which we are all familiar and feel like we know what we mean when talking about it. It is perhaps best defined as a probability, namely “how probable do we think our response was correct”. However, to extract measures of confidence from behavioral data or from neural activity is difficult, because confidence is confounded with other measures of performance in any task or decision. Dr. Kepecs introduces a scheme in which confidence can be extracted as a measure in decision making tasks. In particular, he proposes that average confidence in one’s performance in a task whose parameters can be varied, should increase linearly with task accuracy as those parameters are varied. Here accuracy is defined as the fraction of trials one is correct in the task with those parameters. Furthermore, one can consider trials of low confidence versus high confidence in a task with fixed parameters. Accuracy would be lower in the trials with low confidence, but in both cases as disriminability is increased (meaning the task gets easier) then accuracy increases from chance to perfect performance. Combining these relationships, Dr. Kepecs shows that as discriminability increases and the task gets easier, confidence increases with disciminability in those trials where the response is correct, but perhaps surprisingly, confidence decreases with increased discriminability in those trials where the response is incorrect.
The proposed relationships are borne out in a perceptual decision-making task in which a human indicates whether the rate of “clicks” heard in their left ear is higher or lower than the rate in their right ear, before indicating their confidence in the choice. Similar relationships arise in experiments with rats in odor discrimination tasks whereby the animal’s confidence is inferred by the length of time they are willing to wait for a reward on a trial – if they are highly confident that they are correct and the reward will come, then they will wait longer.
The above relationships allow Dr. Kepecs’ group to identify confidence-selective neurons in the lateral orbitofrontal cortex of rats, a location the group selected because lesion of this brain region impairs the abovementioned behavioral relationships with confidence. Furthermore, if confidence is identified as the probability that a behavior was optimal, then its value is needed in any mechanism for reinforcement learning that involves a reward prediction error. One area of the brain associated with such reinforcement learning is the ventral tegmental area (VTA) and indeed Dr. Kepecs’ group found dopaminergic neurons there whose activity was representative of confidence.
Finally, Dr. Kepecs showed preliminary evidence for a link between deviations from the above-described normal/ ideal relationships between confidence and other task parameters, and propensity for psychiatric disorders such as schizotypy.