Pupil Dynamics Reflect Uncertainty-Driven Adjustments of Probability Learning

This study demonstrates that probability learning in dynamic environments relies on the computationally rational integration of distinct uncertainty factors, where phasic and tonic pupil dilations respectively track change point probability and prior uncertainty to modulate learning rates.

The Gist

The Big Idea: How Our Brains Learn When Things Change

Imagine you are trying to guess the weather. If you live in a place where it rains every Tuesday, you get very confident in that prediction. But what if, suddenly, the weather starts changing unpredictably? One day it's sunny, the next it's a blizzard, and the next it's a heatwave.

To survive in this chaotic world, your brain has to do two things:

1. Realize the rules have changed. (e.g., "Wait, it's not just raining on Tuesdays anymore; the whole system is broken!")
2. Decide how much to trust your old guesses. (e.g., "Should I ignore my old weather map completely, or just tweak it a little?")

This study is about how our brains handle this "uncertainty" and how a tiny muscle in our eyes—the pupil—acts like a window into this mental process.

---

The Experiment: The "Roulette Wheel" Game

The researchers asked people to play a game. Imagine a spinning roulette wheel that is split into two colors: Blue and Yellow.

• The wheel is rigged. Sometimes it's 80% Blue, sometimes 20% Blue.
• The participants had to guess the current ratio (e.g., "I think it's 60% Blue right now").
• The Twist: Every so often, without warning, the wheel's settings change. The "hidden probability" flips.

The participants had to update their guess after every single spin. The researchers wanted to see: How fast do people change their minds when the rules change?

The Two Types of "Uncertainty"

The researchers found that our brains track two different kinds of "doubt" when learning:

1. The "Wait, What?" Factor (Change Point Probability):

   • Analogy: Imagine you are driving down a familiar street, and suddenly a giant pothole appears. You are shocked. You immediately think, "Something big just happened here!"
   • In the brain, this is the sudden realization that the environment has changed. It's a sharp, sudden spike in alertness.

2. The "I'm Not Sure" Factor (Prior Uncertainty):

   • Analogy: Imagine you are trying to guess the price of a rare coin, but you've never seen one before. You have no idea what it's worth. You are generally unsure.
   • In the brain, this is a general feeling of confusion or lack of confidence in your current knowledge. It's a slow, steady fog of doubt.

The Secret Weapon: The Pupil

The researchers used special cameras to watch the participants' pupils. They discovered that the pupils act like a dual-mode dashboard for the brain's learning system:

• The "Flash" (Phasic Pupil Dilation):

  • When the "Wait, What?" factor happens (a sudden change in the game), the pupil flashes open quickly.
  • Metaphor: It's like a camera flash going off. It signals, "Hey! Pay attention! Something new just happened!" This helps the brain reset and learn from the new information immediately.

• The "Glow" (Tonic Pupil Dilation):

  • When the "I'm Not Sure" factor is high (general confusion), the pupil stays wide open for a longer time.
  • Metaphor: It's like a streetlamp that stays on all night. It signals, "We are in a foggy area; keep your sensors open and be ready to learn slowly." This helps the brain gather more data before making a firm decision.

The Big Discovery: Which One Rules?

In previous studies (where people guessed numbers or locations), the sudden "Flash" (Change Point) was the main driver of learning.

But in this study (guessing probabilities), the "Glow" (Prior Uncertainty) was the boss.

• Why? Because guessing probabilities is harder. You can't tell just by looking at one spin if the rules changed. You need to gather a lot of evidence.
• So, the brain relies more on that steady "Glow" of uncertainty. It says, "I'm not confident in my current guess, so I need to be flexible and keep learning."

The "Mediator" Connection

The most exciting part of the paper is that the researchers proved the pupils aren't just watching the brain; they are helping the brain learn.

• They found that the size of the pupil (the "Glow" and the "Flash") actually carries the message from the brain's uncertainty to the brain's learning speed.
• Analogy: Think of the pupil as a messenger. When the brain feels uncertain, it sends a messenger (the pupil) to the learning center to say, "Slow down and be careful!" or "Speed up and pay attention!" The learning rate changes because of this message.

Summary in One Sentence

When we are learning in a confusing world, our pupils act as a dual-signal system: a quick flash tells us to react to sudden changes, while a steady glow tells us to stay open and flexible when we generally don't know what's going on, and these signals directly control how fast we learn.

Technical Summary

1. Problem Statement

Adaptive learning in dynamic environments requires the brain to distinguish between stochasticity (random noise) and volatility (abrupt changes in environmental rules). While previous research has established how the brain adjusts learning rates in magnitude-based inference tasks (e.g., tracking spatial positions or numerical values), the mechanisms governing probability learning (estimating the likelihood of discrete outcomes) remain poorly understood.

In probability learning, individual observations carry less information than in magnitude learning, making change points harder to detect. Normative (Bayesian) theory suggests that in such contexts, prior uncertainty (confidence in current beliefs) should dominate the adjustment of learning rates, rather than change point probability (the likelihood of an abrupt environmental shift). However, the neural mechanisms—specifically the role of arousal systems in encoding these distinct uncertainty factors and regulating learning—were unknown.

2. Methodology

Experimental Design:

• Task: Participants performed a dynamic probability learning task involving a sequence of 75 trials. They were told to estimate the hidden generative probability of a "roulette wheel" (proportion of blue vs. yellow).
• Dynamics: The hidden probability changed abruptly at unpredictable times (change points). The generative probability was sampled uniformly between 0.1 and 0.9.
• Measurement:
  • Behavioral: Participants continuously updated their probability estimates using a semi-circular slide tracker.
  • Physiological: Pupil diameter was recorded at 60 Hz using eye-tracking (Tobii Pro Nano/Fusion) while participants maintained a fixed gaze on a central target to minimize saccadic artifacts.
• Participants: 32 recruited; 31 included in behavioral analysis; 28 included in pupil analysis after quality control.

Computational Modeling:

• Ideal Observer Model: A Bayesian ideal observer model (Hidden Markov implementation) was used to compute the optimal posterior distribution of the hidden probability.
• Latent Variables Extraction: The model extracted trial-by-trial values for:
  1. Apparent Learning Rate ($\alpha_t$): Calculated as the ratio of the belief update to the prediction error.
  2. Prior Uncertainty ($u_t$): The standard deviation of the prior belief distribution (epistemic uncertainty).
  3. Change Point Probability ($\Omega_t$): The probability that the hidden state changed since the last observation.

Data Analysis:

• Regression Analysis: Multiple linear regression was applied to pupil diameter signals across a peri-stimulus window (-250 ms to 1400 ms).
  • Phasic Signal: Baseline-corrected pupil dilation (relative to -250 to 0 ms).
  • Tonic Signal: Absolute (non-baseline corrected) pupil diameter.
• Mediation Analysis: Used to test whether pupil signals mediated the relationship between latent uncertainty factors and apparent learning rates.
• Lag Optimization: An algorithm was employed to shift participant responses by one trial to account for potential temporal lags in belief updating, ensuring accurate alignment with model-derived variables.
• Validation: A secondary analysis was performed on an auditory probability learning task (n=15) to rule out motor artifacts from the visual slider task.

3. Key Contributions

1. Dissociation of Uncertainty Factors in Probability Learning: The study confirms that in probability learning, prior uncertainty is the dominant driver of learning rate adjustments, superseding change point probability. This contrasts with magnitude learning tasks where change point probability is often the primary driver.
2. Distinct Arousal Signatures: The paper provides the first evidence that phasic and tonic pupil responses encode distinct latent uncertainty factors:
   • Phasic Pupil Dilation: Tracks change point probability (peaking ~630ms post-stimulus).
   • Tonic Pupil Dilation: Tracks prior uncertainty (present throughout the epoch, including pre-stimulus).
3. Mechanistic Link to Learning: Through mediation analysis, the study demonstrates that these pupil signals are not merely correlates but mediate the effect of uncertainty on learning behavior. Specifically, tonic pupil size partially explains how prior uncertainty influences the learning rate.
4. Amodal Generalizability: The dissociation between phasic/tonic signals and uncertainty factors was replicated in an auditory task, confirming that these mechanisms are intrinsic to the learning process rather than artifacts of visual-motor responses.

4. Key Results

• Behavioral Performance: Participants closely followed the optimal Bayesian solution (Pearson's $r = 0.81$). Their apparent learning rates increased significantly after change points, mirroring normative theory.
• Drivers of Learning:
  • Prior Uncertainty was a significant positive predictor of apparent learning rates ($\beta = 0.09$).
  • Change Point Probability initially showed no significant direct effect on learning rates in raw data. However, after correcting for trial-wise lags (where participants sometimes updated beliefs one trial late), a significant relationship emerged, though prior uncertainty remained the dominant factor.
• Pupil Dynamics:
  • Phasic Response: Showed a significant effect of change point probability starting 630ms after stimulus onset. This aligns with the cognitive processing of surprise or prediction errors.
  • Tonic Response: Showed a significant effect of prior uncertainty across the entire epoch, including before stimulus onset. This suggests a sustained state of arousal reflecting the current level of belief uncertainty.
• Mediation Effects:
  • Tonic Pupil: Significantly mediated the effect of prior uncertainty on learning rates (Indirect effect significant).
  • Phasic Pupil: Significantly mediated the effect of change point probability on learning rates, even though the direct effect of change point probability on learning was weak or non-significant in some models. This suggests the pupil signal captures the mechanism of regulation even when the behavioral output is noisy.

5. Significance and Implications

• Neurocomputational Mechanism: The findings support the hypothesis that the Locus Coeruleus-Norepinephrine (LC-NA) system regulates learning through two distinct modes:
  • Tonic Mode: Reflects global uncertainty (prior uncertainty), modulating neural gain to facilitate exploration and belief updating when confidence is low.
  • Phasic Mode: Reflects specific events (change points/surprise), triggering rapid updates to reset beliefs.
• Context-Dependent Learning: The study highlights that the brain adapts its learning strategy based on the task domain. In probability learning (low information gain per trial), the system relies heavily on the state of uncertainty (tonic) rather than just the detection of change (phasic).
• Physiological Biomarker: Pupilometry is validated as a non-invasive tool to dissociate and monitor distinct computational variables (uncertainty vs. surprise) in real-time, offering a window into the "confidence-weighting" mechanisms of the human brain.
• Clinical Relevance: Understanding how arousal systems regulate learning in volatile environments provides a framework for investigating disorders characterized by maladaptive learning (e.g., anxiety, OCD, or schizophrenia), where the balance between exploration and exploitation may be disrupted.

In summary, this paper establishes that human probability learning is governed by a computationally rational integration of latent uncertainty factors, implemented via distinct phasic and tonic arousal responses that can be tracked non-invasively through pupil dynamics.