🧠 neuroscience

Dissociable Microstructural Correlates of Learning Rate and Learning Noise in Gamified Reward-Based Decision-Making

This large-scale study combining quantitative MRI and computational modeling reveals that individual differences in reward-based decision-making are driven by distinct microstructural correlates, where cerebellar myelination predicts learning rate while precentral gyrus myelination and iron concentration determine learning noise.

Original authors: Vejloe, M., Nikolova, N., Banellis, L., Tyrer, A., Skvortsova, V., Hauser, T. U., Allen, M.

Published 2026-02-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vejloe, M., Nikolova, N., Banellis, L., Tyrer, A., Skvortsova, V., Hauser, T. U., Allen, M.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine your brain is a bustling, high-tech city where millions of decisions are made every second. Some people navigate this city with a GPS that updates instantly and perfectly, while others have a GPS that is a bit glitchy, sometimes taking a wrong turn or hesitating before making a choice.

This study is like a deep-dive inspection of the roads and power lines inside that city to see why some people's "decision GPS" works better than others.

Here is the breakdown of what the researchers found, using simple analogies:

1. The Experiment: A Game of Space Cows

The researchers didn't just ask people to think; they made them play a video game.

The Setup: Imagine two "Space Cows" (one brown, one black-and-white) standing in front of you. You have to pick one to get "space milk" (points).
The Catch: The cows don't give points randomly. Sometimes the brown cow is generous, sometimes the black-and-white one is. The rules change slowly over time.
The Goal: You have to figure out which cow is currently paying out the most and stick with it, but you also need to be ready to switch if the rules change.

The researchers used a computer model to measure two specific things about how you played:

Learning Rate: How fast you figured out which cow was the "good" one. (Fast learner vs. slow learner).
Learning Noise: How "jittery" or inconsistent your updates were. Even if you knew the rules, did your brain add a little bit of static or confusion to the signal? (Stable learner vs. noisy learner).

2. The Scan: Looking at the Brain's "Infrastructure"

Instead of just looking at which parts of the brain lit up (like seeing which lights are on in a city), the researchers used a special MRI scanner to look at the physical building materials of the brain. They focused on two things:

Myelin (The Insulation): Think of this as the plastic coating on an electrical wire. More myelin means the signal travels faster and cleaner.
Iron (The Fuel/Power): Iron is crucial for the brain's energy systems, particularly for dopamine (the "reward chemical").

3. The Big Discovery: Two Different Roads for Two Different Problems

The study found that "Learning Rate" and "Learning Noise" are not the same thing. They are controlled by different parts of the brain's infrastructure.

🚀 The "Speed" Road: The Cerebellum

What they found: People who learned the game rules faster had thicker, better-insulated wires (more myelin) in a part of the brain called the Cerebellum.
The Analogy: The Cerebellum used to be thought of only as the "balance and movement" center (like a tightrope walker). But this study shows it's also the brain's high-speed data hub. If the insulation here is thick, the brain can process "Hey, that cow gave me points!" and update its strategy instantly.

📡 The "Static" Road: The Motor Cortex

What they found: People who had noisy learning (where their brain added confusion or jitter to the updates) had more myelin and iron in the Precentral Gyrus (the Motor Cortex).
The Analogy: This is the part of the brain that plans your physical movements (like pressing a button). The researchers suggest that if the "wiring" here is too thick or has too much iron, it might be like a radio station with too much signal strength—it creates static.
- Even though this is a thinking game, you have to physically press a button to choose. The study suggests that the "noise" in your decision-making might actually be "static" leaking in from the part of your brain that is planning your hand movement.

4. Why Does This Matter?

Think of your brain like a computer.

Old View: We used to think all learning problems were just one big "glitch" in the software.
New View: This study shows that the "glitch" could be a hardware issue in the Cerebellum (making you slow to learn) OR a hardware issue in the Motor Cortex (making your updates jittery).

The Real-World Impact:
This helps us understand why some people struggle with conditions like ADHD or OCD.

If someone has trouble sticking to a routine because they are "noisy" in their decisions, maybe their "Motor Cortex wiring" is the culprit.
If someone just can't learn new rewards quickly, maybe their "Cerebellum insulation" needs attention.

Summary

Fast Learners = Have better "insulation" (myelin) in the Cerebellum (the brain's data hub).
Noisy/Jittery Learners = Have specific "wiring" and "iron" patterns in the Motor Cortex (the brain's movement planner).

The brain isn't a single machine; it's a complex city where different neighborhoods handle different parts of the learning process. By understanding the physical roads and wires, we can finally start to understand why we all learn differently.

1. Problem Statement

While humans universally learn from rewards through trial and error, there is substantial individual variability in the speed (learning rate) and precision (learning noise) of this process. Traditional reinforcement learning (RL) models often attribute behavioral variability solely to the choice-selection stage (stochasticity), assuming the learning update process itself is deterministic. However, recent computational theories suggest that learning itself is subject to noise.

The biological origins of these individual differences in learning efficiency and noise remain poorly understood. Specifically, it is unclear how brain microstructure—particularly variations in myelination and cortical iron concentration—relates to the computational parameters of reward learning. Previous studies have mapped these variables to functional activations (fMRI), but structural correlates using quantitative MRI (qMRI) are underexplored.

2. Methodology

Participants and Task

Cohort: Data was drawn from the "Visceral Mind Project" at Aarhus University. After rigorous quality control (QC) and exclusion for motion artifacts or medical issues, the final sample consisted of 239 participants for myelin analysis (R1) and 232 for iron analysis (R2*).
Task: Participants performed a gamified two-armed bandit task ("Milky Way game") involving 72 trials. They chose between two "space cows" with drifting reward probabilities to accumulate points.
Computational Modeling: A Noisy Learning Model was fitted to individual choice data. Key parameters extracted included:
- Learning Rate ( $\alpha$ ): Scales the update of value based on prediction errors.
- Learning Noise ( $\zeta$ ): An additive noise term in the value update equation, scaling with the magnitude of the prediction error (Weber's law).
- Choice Stochasticity ( $\beta$ ): Temperature parameter for the softmax action selection policy.

Neuroimaging and Analysis

Acquisition: High-resolution (0.8 mm³) quantitative MRI (qMRI) was performed on a 3T Siemens Prisma scanner using a Multi-Parameter Mapping (MPM) protocol.
Quantitative Maps:
- R1 (Longitudinal Relaxation Rate): A proxy for myelination.
- R2 (Effective Transverse Relaxation Rate):* A proxy for tissue iron concentration.
- MT (Magnetization Transfer): A proxy for myeloarchitectural integrity.
Statistical Approach: Voxel-Based Quantification (VBQ) was employed. Unlike standard Voxel-Based Morphometry (VBM), VBQ preserves quantitative data without modulation for volume, minimizing partial volume effects.
- Analysis: Whole-brain multiple linear regression was conducted in SPM12.
- Covariates: Age, gender, BMI, and Total Intracranial Volume (TIV) were included as nuisance regressors.
- Correction: Significance was determined using Threshold-Free Cluster Enhancement (TFCE) with 5,000 permutations, corrected for Family-Wise Error (FWE) at $p < .05$ .

3. Key Contributions

Dissociation of Learning Components: The study provides the first evidence that learning rate and learning noise are supported by distinct, dissociable neurobiological pathways rather than a unitary process.
Microstructural Specificity: It moves beyond functional activation to identify specific microstructural correlates (myelin density and iron concentration) underlying computational learning parameters.
Novel Regional Findings: It identifies the cerebellum and motor cortex as critical hubs for reward learning variability, challenging the view that these regions are solely motor-related.

4. Key Results

A. Learning Rate and the Cerebellum

Finding: Increased R1 values (myelination) in the right cerebellum (exterior) were positively associated with a higher learning rate for the chosen object.
Interpretation: This suggests that the structural integrity of the cerebellum facilitates the speed at which reward prediction errors are integrated into value estimates. This aligns with emerging views of the cerebellum as a hub for cognitive and reward processing, not just motor control.

B. Learning Noise and the Motor Cortex/Cerebellum

Finding 1 (Cerebellum): Increased R1 values (myelination) in the right cerebellum were also positively associated with learning noise.
Finding 2 (Motor Cortex): Increased learning noise was associated with both:
- Higher R2 values (iron concentration)* in the left medial precentral gyrus.
- Higher R1 values (myelination) in the same left medial precentral gyrus region.
Interpretation: The overlap of iron and myelin in the motor cortex suggests that the precision of value updating is linked to the microstructural properties of the motor system. The authors hypothesize that learning noise (imprecision in value updates) may propagate into the motor planning and selection systems, or that the motor cortex actively integrates expected action values.
Null Result: No significant associations were found between choice stochasticity ( $\beta$ ) and any microstructural markers, reinforcing the idea that learning noise and choice stochasticity are distinct computational phenomena.

5. Significance and Implications

Theoretical Impact: The findings support the hypothesis that learning itself is a noisy process governed by specific neural substrates. The dissociation between learning rate (cerebellar myelin) and learning noise (cerebellar myelin + motor cortical iron/myelin) implies that "how fast" one learns and "how precisely" one learns are biologically distinct traits.
Clinical Relevance: The study highlights potential biomarkers for psychiatric disorders characterized by reward processing deficits, such as ADHD (often linked to high learning noise) and OCD. Future research could investigate whether microstructural variations in the cerebellum and motor cortex underlie the "non-greedy" decision-making patterns seen in these populations.
Methodological Advancement: By combining computational modeling with quantitative MRI (VBQ), the study demonstrates a robust framework for linking specific behavioral parameters to biophysically meaningful tissue properties (myelin and iron), offering a more precise alternative to functional imaging alone.
Future Directions: The authors suggest that the link between iron/myelin and learning may be influenced by nutritional status or aging. They propose future longitudinal studies and multimodal imaging (combining qMRI with PET to measure dopamine receptor density) to establish causality and clarify the role of the dopaminergic system.