If you are asked to estimate your confidence in something—such as a decision you just made—you can readily answer. What is the neural basis for such judgments? Is knowledge about beliefs an example of the human brain's capacity for self-awareness? Or is there a simpler explanation that might suggest a more fundamental role for confidence in brain function across species?
When we started our studies, these questions seemed to us to belong to the realm of the philosophical. As we soon discovered, this has also been a contentious issue in the psychological literature, because confidence has often been thought of as an instance of "metacognition"—thinking about thinking—that requires an advanced neural architecture available only in primate brains (1–3). Because we study rodents, we wondered: Could rats also determine how sure they are about a decision they’ve just made? It turns out the answer is yes! And soon we also discovered neurons that tracked how confident rats were (4). Initially my colleagues and I were interested in the principles of decision-making and a brain area called the orbitofrontal cortex (OFC), whose damage impairs decision-making in humans (5). Because rats have an excellent sense of smell, we trained them in a simple decision task to determine the dominant odor in a two odor mixture. For each correct decision, they received a drop of water. Critically, rats were trained to nearly perfect accuracy for pure odors but their performance dropped as the odor mixtures got more difficult to classify. Therefore we could "dial in" the difficulty of individual decisions by varying the odor mixture ratio. Importantly, our rats could perform hundreds of trials each day, allowing statistical analysis to relate trial-to-trial behavioral observations to neural activity.
Fig. 1. Neural tuning curves for stimuli, choices, and confidence.
(A and B) Hypothetical tuning curves for an odor mixture and a choice-coding neuron. Firing rate is a function of the odor mixture ratio for the odor-coding neuron, independent of choice. Firing rate is a function of choice for the choice-coding neuron, independent of odor mixture ratio. (C) Stimulus and outcome of selective firing of an example neuron from the OFC, (D) Performance accuracy plotted as a function of the firing rate for the same neuron as in (C). (E and F) Model predictions for decision uncertainty. Shown are mean uncertainty estimates generated by the model as a function of stimulus and trial outcome, and mean accuracy of model choices as a function of uncertainty. Maximal uncertainty (or equivalently minimal confidence) implies chance-level performance; 50% in this two-alternative choice task.
In order to understand the neural basis of rats’ decisions, we recorded the activity of neurons in the OFC as the rats were performing the decision task. How could we tell which messages the neural activity represented? The standard answer is to construct neural tuning curves (the firing rate of neurons plotted against a stimulus feature) to determine which stimuli neurons care about (i.e., which ones change their firing rate), and which features they are insensitive to (no change in firing rate). For instance, in a sensory area, one might expect neurons to fire as a function of the odor stimulus [the odor mixture ratio, for example (Fig. 1A)] independent of what the animal chose, whereas in a motor area, one would expect neurons to fire as a function of the response [left or right choices (Fig. 1B)], regardless of the stimulus that determined the animal’s choice. Therefore, we examined how the firing of the neurons changed as a function of the odor stimulus (the odor mixture ratio) and the final outcome (right or wrong answer). The results were a surprise: Some neurons varied their firing as a function of both the stimulus and the outcome in a peculiar fashion (Fig. 1C). The general observation that both stimulus and outcome mattered wasn’t entirely surprising when studying OFC, a brain region far removed from the sensory or motor peripheries. But the particular tuning curves we obtained called for a better explanation. We suspected that these neurons represented decision confidence, or its inverse, decision uncertainty. But how could we determine this, given that confidence is an internal variable that cannot be directly accessed? Employing an emerging paradigm designed to address this (6–9), we constructed quantitative models in terms of controlled and observable behavior, in our case the stimuli and the outcomes, for the unobservable internal variable of "confidence." Using a simple model of choice behavior, we computed an estimate of confidence for each choice. This approach yielded quite specific predictions about what a representation of confidence would look like based on a few relatively general assumptions (Fig. 1, E and F). It turned out that neural activity recorded from a large population of OFC neurons was consistent with this model. Moreover, these neurons’ firing tracked the difficulty of decisions and, as expected for a good confidence estimate, could even predict the outcomes of decisions. For instance, when these neurons fired at high rates, they predicted near-chance performance by the animal, whereas low rates predicted high performance (Fig 1D).
One may question whether these neurons might represent "anxiety," "arousal," or "attention," instead of "confidence." Of course they might. But rather than labeling such neural signals with psychological concepts, what we attempted to determine was what computation generated these signals. Therefore our label "confidence" is based on howa particular representation is computed and not whatfunction it ultimately serves. In other words, confidence is a label for a variable in the model that conforms to both the formal and common-sense notions of decision confidence, and not an anthropomorphic description of a behavioral function. Most importantly, models are concrete: They can be tested, disproved, and iteratively improved, moving the scientific debate forward—unlike arguments about psychological labels.
If we truly discovered confidence-signaling neurons, then our rats should be able to perform confidence judgments. To test this, we modified the decision task by delaying reward delivery and allowing rats to abort and restart trials during the delay. In this new version of the task, the decision whether to abort or not should depend on the degree of confidence: It was best to stay and wait for reward when certain and leave and start a new trial when uncertain. Indeed, we found that the pattern of trial restarts reflected the predictions of the same model of decision confidence that also captured the neural data, demonstrating that rats are able to behaviorally report their confidence in a decision.
Taken together, our results demonstrate that confidence judgments need not involve mysterious acts of self-awareness, but instead can be based on a simple computation performed by generic neural circuits (10, 11) and are not limited to primates (12). Of course, our results are just a first step towards reducing the fuzzy notion of confidence to a well-formulated problem that neuroscience methods can address, but I propose that estimating confidence will turn out to be fundamental to animal behavior and for computations in the nervous system.