Universal Coding for Uncertainty?

To make decisions in a fluctuating environment, it can be helpful to anticipate upcoming events by estimating their likelihood. For highly probable or improbable events, we can be fairly certain about our estimates. However, events are often more stochastic, and hence our likelihood estimations become uncertain. Encoding of such uncertainty is particularly motivationally relevant when the events are related to desirable or detrimental outcomes (e.g., “is this road the fastest route to my destination right now?”). In contrast, the encoding of uncertainty of neutral events is less motivationally relevant; but it can still be useful if it provides general knowledge about the environment (e.g., “is the road generally busy?”). These two forms of uncertainty are called valenced uncertainty and generic uncertainty. Despite their qualitative difference, both forms of uncertainty provide adaptive value for reacting to the environment.

A fundamental question is whether a universal coding mechanism is used to represent all forms of uncertainty or if instead uncertainty is represented differently depending on the motivational value. The human brain has widely been reported to use a universal coding mechanism in valuation systems. The neural common currency hypothesis suggests that for choices of all kinds, even qualitatively different (e.g., money, food, goods), goodness or utility is encoded in the same ventromedial prefrontal cortex (vmPFC) region; this hypothesis has received support from multiple lines of research (Levy and Glimcher, 2012). Since value representations are closely tied to uncertainty in decision-making, it is possible that a universal coding mechanism also applies to the encoding of different forms of uncertainty. On the other hand, other studies have revealed that valenced words are better memorized and recalled than neutral words (Megalakaki et al., 2019). This suggests that valenced and neutral information are processed differently. Moreover, a Darwinian perspective suggests the brain should evolve mechanisms that are optimized for stimuli of greater survival value. Valenced stimuli, compared with neutral stimuli, require prompt responses to obtain appetitive outcomes or avoid aversive outcomes, and hence it is likely that they involve distinct neural processing. Accordingly, the encoding of valenced uncertainty and generic uncertainty may use different mechanisms. However, formal testing of the universal coding versus multimechanism hypotheses for uncertainty encoding has been lacking, because most previous studies on uncertainty encoding focused only on valenced uncertainty.

A recent study by Kim et al. (2024) has shed some light on the neural mechanisms underlying the encoding of uncertainty by introducing a neutral outcome to dissociate valenced uncertainty from generic uncertainty. They designed a pavlovian conditioning task where participants learned to associate different visual cues with the delivery of different kinds of liquid. The liquids were either valenced (appetitive or aversive) or unvalenced (neutral). Crucially, each visual cue informed participants that a specific liquid would be delivered with 0, 50, or 100% probability. Thus, the outcome was certain (uncertainty was low) for cues associated with 0 or 100% probability of delivery, whereas uncertainty was high for cues associated with 50% probability of delivery. Moreover, uncertainty was valenced for cues signaling probabilistic delivery of aversive or appetitive liquid and unvalenced for cues signaling probabilistic delivery of neutral liquid. After learning the contingency between the visual cues and the delivery of different liquids, participants underwent functional magnetic resonance imaging (fMRI) during which they were presented with the learned visual cues and asked either to rate the pleasantness ratings of the cues or to report the cue contingency. Associated liquid was then delivered according to the learned probabilities after responses.

Kim et al. (2024) first tested whether there are any differences between participants’ behavioral responses to valenced and neutral cues. In line with previous findings, participants responded faster to valenced cues. However, linear regression of participants’ reaction times (RTs) against valenced uncertainty and generic uncertainty revealed no significant effect, suggesting that participants’ faster responses to valenced cues was not dependent on uncertainty. To inspect these results more closely, a domain-specific analysis was performed to break down the uncertainty into the uncertainty of appetitive, aversive, and neutral cues. This showed that participants responded more quickly with higher uncertainty of appetitive cues and more slowly with higher uncertainty of aversive cues, while uncertainty of neutral cues had no significant effect. These behavioral results suggest that valenced uncertainty and generic uncertainty may involve different underlying mechanisms.

Consistent with the behavioral results, the neural data analysis provided support for different mechanisms underlying uncertainty encoding. A general linear model revealed a dissociation between the neural correlates of valenced uncertainty and generic uncertainty. Neural correlates of valenced uncertainty were found in the dorsal anterior insula and other regions that are commonly involved in value-based decision-making, such as the lateral orbitofrontal cortex and lateral frontopolar cortex (FPl). On the other hand, activity in regions in the occipital lobe, such as the calcarine and lingual gyrus, was correlated with generic uncertainty. This dissociation in neural correlates of valenced uncertainty and generic uncertainty is in line with the multimechanism hypothesis—valenced and generic uncertainty are processed by separate underlying mechanisms.

One of the notable features of Kim and colleagues’ study is the use of a parsimonious experimental design to disentangle the neural signals of different forms of uncertainty. Specifically, they used minimal levels of probability to dissociate uncertainty from probability and directly tested the valenced and generic uncertainty signals without requiring uncertainty-dependent responses during the fMRI task. However, neural correlates of uncertainty were observed in multiple regions, raising the possibility that the actual regions underlying uncertainty encoding were confounded with other regions that underlie related processes. For example, Kim and colleagues found that the activity in the insula is one of the neural correlates of valenced uncertainty and activity was particularly high with aversive cues. Because the insula is a key region in affect processing, particularly in negative affects (Berntson et al., 2011), insula activity here may have stemmed from negative emotions induced by uncertainty rather than the encoding of uncertainty per se. Experimental designs that are dedicated to distinguish uncertainty from other related processes are warranted in future studies to provide convergent evidence of the neural correlates of uncertainty.

Another limitation of the work by Kim et al. (2024) is that their neural data analysis focused on the ventral brain, while previous findings suggested the dorsal anterior cingulate cortex (dACC) is a strong candidate for encoding both generic uncertainty and valenced uncertainty. DACC's capacity for universal coding has been suggested by its engagement in encoding uncertainty across qualitatively different tasks. In a decision–redecision paradigm, participants were asked to report their subjective uncertainty about their decisions and then were allowed to rectify the decisions (Su et al., 2022). DACC showed an uncertainty signal at the redecision phase, and, crucially, this uncertainty signal was consistently observed across different tasks, including perceptual decision-making, recognition memory, and a Sudoku-problem task. In reinforcement learning, where the options were associated only with reward (Trudel et al., 2021) or with both reward and loss (Payzan-LeNestour et al., 2013), an uncertainty signal was also found in dACC. On the other hand, in a decision-making task that required estimating the composition of two nonvalenced decks of cards through multiple trials of observation, dACC showed activity correlated with the subjective uncertainty (Stern et al., 2010). Because Kim et al. (2024) focused only on the ventral brain, whether dACC shows a universal coding signal of uncertainty remains unclear and warrants further testing. This can be tested by extending the current univariate analysis to the dorsal brain and further supplemented with a multivoxel pattern analysis to examine the activity patterns across multiple voxels.

In light of the findings by Kim et al., an intriguing question that arises is whether the brain regions showing uncertainty signals also engage in using the uncertainty for guiding behavior. For example, Kim et al. (2024) found neural correlates of valenced uncertainty in the FPl, which has been previously found to engage in guiding value-based decision-making (Law et al., 2023). However, the connection between the identified neural correlates of uncertainty and decision-making could not be affirmed unless additional testing is performed. In a related context, subjective value is often considered a key attribute in guiding behavior, with its neural correlates commonly observed in the vmPFC and the posterior cingulate cortex (PCC) activities. By evaluating choice-relevant and choice-irrelevant option attributes in a decision-making task, Grueschow et al. (2015) showed that only the vmPFC has a role in making choices. Similarly, a paradigm that includes responses depending on valenced uncertainty or generic uncertainty can help scrutinize which brain regions that encode uncertainty are pertaining to guiding adaptive behavior.

All in all, Kim et al. (2024) provided a framework for dissociating valenced uncertainty and generic uncertainty, which enabled testing the multimechanism hypothesis versus the conventional universal coding hypothesis for encoding of uncertainty. The foundational work by Kim et al. (2024) can now be extended by making the behavioral responses uncertainty-dependent, such that the encoding of uncertainty could be examined in a neurobehavioral manner, thereby providing a clearer picture of the underlying neural substrates. Because of the close ties between uncertainty encoding and other areas in cognitive neuroscience, such as decision-making, information sampling, and working memory, further studies to deepen our understanding of the neural mechanisms of uncertainty encoding are warranted.