Task-evoked pupil responses reflect internal belief states

Olympia Colizoli, Jan Willem de Gee, Anne E Urai, Tobias H. Donner

TL;DR

This study reveals that pupil dilation, a physiological marker of arousal, reflects internal cognitive states like decision confidence and prediction error during perceptual tasks. By analyzing pupil responses before and after receiving feedback in a challenging choice task, the researchers found that pupil dilation tracks these internal belief states, offering a non-invasive window into how the brain processes uncertainty and learning.

What Problem Does This Paper Address?

Making decisions in the real world often involves navigating uncertainty. We don't always have complete information, and our confidence in a decision can vary. How does the brain manage this uncertainty, and how does it learn from outcomes? Previous research has suggested that the brain's arousal systems, which can be measured peripherally through pupil dilation, play a role in signaling uncertainty. However, it was unclear whether these arousal signals specifically reflected internal computational variables related to decision-making and learning, or just general task engagement. This paper aims to bridge this gap by investigating if task-evoked pupil responses can specifically track two key computational variables: decision confidence (how sure we are about our choice) before feedback, and prediction error (the difference between expected and actual reward) after feedback.

Methods

The researchers designed a challenging two-alternative forced-choice (2AFC) random dot motion discrimination task. Participants had to decide whether the dots were moving up or down. The difficulty of the task was manipulated by varying the coherence (strength) of the motion. Crucially, feedback about the correctness of their choice and a monetary reward was delayed, allowing for independent measurement of pupil responses related to decision formation and feedback processing.

Pupil diameter was continuously monitored using an EyeLink 1000 system. To analyze the pupil responses, they employed a "sustained time window" approach, focusing on pupil dilation that persisted long after the initial stimulus or response, suggesting a cognitive rather than purely sensory origin. They compared the observed pupil responses to predictions from two computational models:

  • Belief State Model: This model assumes that internal belief states (like decision confidence based on the observer's internal decision variable) are used to predict rewards and compute prediction errors.
  • Stimulus State Model: This model assumes that these variables are based solely on external stimulus information, not internal beliefs.

Statistical analyses, including cluster-based permutation tests and ANOVAs, were used to assess the significance of the findings. Control analyses were performed to rule out confounds from reaction time (RT) and motion energy.

Datasets

The study used data collected from 15 healthy human participants. No new datasets were introduced; the data was collected specifically for this experiment.

Key Insights and Ideas

  • Pupil Dilation as a Window into Internal Beliefs: The most striking insight is that pupil dilation isn't just a general indicator of cognitive load, but can specifically reflect abstract internal computational states like decision confidence and prediction error.
  • Two Distinct Phases of Arousal Signaling: The study identified two key moments where arousal (pupil dilation) is modulated by internal states:
  • Before Feedback (Reward Anticipation): Pupil dilation scales with decision uncertainty (the inverse of confidence). Higher uncertainty leads to greater pupil dilation, suggesting the brain is preparing for potential reward or error.
  • After Feedback (Reward Processing): Pupil dilation reflects prediction errors. Larger pupil responses occur when the outcome (reward or error) deviates from what was expected, particularly when this deviation is influenced by the observer's internal belief about their choice accuracy.
  • Belief State Model Outperforms Stimulus State Model: The experimental results strongly favored the "Belief State Model," indicating that internal subjective states, rather than just objective stimulus properties, are crucial for modulating arousal responses related to decision-making and learning.
  • Sustained Arousal: The pupil responses were sustained, lasting for several seconds after the relevant event (choice or feedback). This suggests that these cognitive signals contribute to ongoing arousal levels, not just transient bursts.

Key Results

  • Pre-feedback: Pupil responses were negatively scaled with decision confidence (positively with uncertainty). This means participants showed larger pupil dilations when they were less confident in their choice.
  • Post-feedback: Pupil responses scaled with uncertainty-dependent prediction errors. This indicates that the brain's arousal system responds not just to the raw outcome, but to how surprising that outcome was given the observer's internal belief about their choice.
  • Interaction Effects: Both pre- and post-feedback pupil responses showed significant interaction effects between decision difficulty and accuracy, consistent with the predictions of the Belief State Model and not the Stimulus State Model.
  • Model Correlation: Group-level correlations showed that pupil responses were significantly more similar to the predictions of the Belief State Model than the Stimulus State Model, both for pre- and post-feedback intervals.

Limitations and Future Work

  • Confounding Factors: While control analyses for RT and motion energy were performed, the authors acknowledge that trial-to-trial variations in these factors can influence pupil responses. However, their key findings remained robust even after controlling for these.
  • Learning vs. Anticipation: The current task focused on reward anticipation and processing rather than explicit learning. Future work should investigate the role of pupil-linked arousal in actual learning processes.
  • Subjective Confidence: The study used a computational model of statistical decision confidence. Future research could explore the relationship between these pupil responses and subjective reports of confidence.
  • Dopaminergic Link: The findings align with patterns observed in midbrain dopamine neurons. Future invasive studies could directly investigate the link between pupil dilation and dopaminergic activity in signaling belief-modulated prediction errors.
  • Generalizability: The study used a specific perceptual task. Further research could explore these effects in more complex decision-making scenarios and across different domains.

Summary updated

Abstract

Perceptual decisions about the state of the environment are often made in the face of uncertain evidence. Internal uncertainty signals are considered important regulators of learning and decision-making. A growing body of work has implicated the brain's arousal systems in uncertainty signaling. Here, we found that two specific computational variables, postulated by recent theoretical work, evoke boosts of arousal at different times during a perceptual decision: decision confidence (the observer's internally estimated probability that a choice was correct given the evidence) before feedback, and prediction errors (deviations from expected reward) after feedback. We monitored pupil diameter, a peripheral marker of central arousal state, while subjects performed a challenging perceptual choice task with a delayed monetary reward. We quantified evoked pupil responses during decision formation and after reward-linked feedback. During both intervals, decision difficulty and accuracy had interacting effects on pupil responses. Pupil responses negatively scaled with decision confidence prior to feedback and scaled with uncertainty-dependent prediction errors after feedback. This pattern of pupil responses during both intervals was in line with a model using the observer's graded belief about choice accuracy to anticipate rewards and compute prediction errors. We conclude that pupil-linked arousal systems are modulated by internal belief states.
BibTeX
@inproceedings{colizoli2018task,
  author = {Olympia Colizoli and Jan Willem de Gee and Anne E Urai and Tobias H. Donner},
  title = {{Task-evoked pupil responses reflect internal belief states}},
  booktitle = {Scientific},
  year = {2018},
  doi = {10.1038/s41598-018-31985-3},
}
Figure