Efficient stabilization of imprecise statistical inference through conditional belief updating

Julie Drevet, Jan Drugowitsch, Valentin Wyart

Abstract

Statistical inference is the optimal process for forming and maintaining accurate beliefs about uncertain environments. However, human inference comes with costs due to its associated biases and limited precision. Indeed, biased or imprecise inference can trigger variable beliefs and unwarranted changes in behavior. Here, by studying decisions in a sequential categorization task based on noisy visual stimuli, we obtained converging evidence that humans reduce the variability of their beliefs by updating them only when the reliability of incoming sensory information is judged as sufficiently strong. Instead of integrating the evidence provided by all stimuli, participants actively discarded as much as a third of stimuli. This conditional belief updating strategy shows good test-retest reliability, correlates with perceptual confidence, and explains human behavior better than previously described strategies. This seemingly suboptimal strategy not only reduces the costs of imprecise computations, but counter-intuitively increases the accuracy of resulting decisions.

Introduction

Statistical inference—the process of updating beliefs based on new data—is mathematically optimal in a vacuum. However, the human brain is a "noisy" machine. We face two major hurdles: sensory noise (the world is blurry) and computation noise (our internal logic is imprecise).

In volatile environments where the "ground truth" changes over time (e.g., a stock market shift or a changing weather pattern), these internal noises are dangerous. If you update your belief based on a weak, noisy signal, you might trigger an "unwarranted change of mind," abandoning a correct belief for a wrong one simply because of a glitch in your internal processing. The authors sought to understand how humans stabilize their beliefs against this internal turbulence.

Methods and Task Design

The researchers used a sequential categorization task called the "Marble Task." Participants saw a series of spatially scrambled marbles (light or dark) and had to decide which of two bags (a "light" bag or a "dark" bag) the marbles were being drawn from.

Experimental task design and behavioral performance

Crucially, the task involved:

  1. Volatility: The bag being used would occasionally switch without warning.
  2. Titration: An adaptive procedure ensured each participant perceived the marbles with a specific level of difficulty (e.g., 80% accuracy), standardizing sensory noise across the group.
  3. Computational Modeling: Human behavior was compared against five different "stabilization" strategies, including perceptual bias, inference lapses, and repetition bias.

The study relied on original behavioral data from 60 healthy adult participants across two experiments. Experiment 1 focused on core categorization and model fitting, while Experiment 2 introduced varying stimulus difficulty and confidence reports.

The Conditional Inference Model

The winning strategy, Conditional Inference, proposes that humans update their beliefs only when the sensory signal exceeds a reliability threshold ($\delta$). This "metacognitive gating" allows the brain to filter out internal computational noise by ignoring unreliable data.

Computational framework of the conditional belief updating model

This strategy offers a unique balance between stability and flexibility. Unlike other biases (like repetition bias) that slow down reactions to environmental changes, conditional inference reduces "flip-flopping" without significantly hindering the detection of real reversals.

Results and the "Less is More" Paradox

The Conditional Inference model outperformed all other candidates, explaining human behavior with an exceedance probability of > 0.999. Participants discarded an average of 32.5% of stimuli as "unreliable."

Paradoxically, in the presence of computation noise, processing more data can actually decrease accuracy. The authors found that implementing this "suboptimal" discarding strategy actually increased decision accuracy by approximately 3.2% to 5.5% compared to a model that tried to process every stimulus. There is an "optimal" amount of information to ignore that maximizes performance by preventing the propagation of internal noise.

Furthermore, the reliability threshold ($\delta$) showed high test-retest reliability across different days, suggesting it is a stable individual "trait." Individual thresholds also correlated strongly with perceptual confidence ($r = 0.392, p = 0.033$); people who are naturally more confident in their perceptions tend to have lower thresholds and discard less information.

Discussion and Future Work

While the model describes the mathematical process of conditional updating, the underlying neural mechanisms for this "gating" remain to be identified. Future research is needed to determine if this strategy applies to other forms of uncertainty, such as category ambiguity (clear data that is a poor predictor of outcomes).

Given its stability as a trait, the reliability threshold could serve as a useful metric for studying psychiatric conditions—such as OCD or schizophrenia—where belief stability and confidence are often disrupted. This work highlights how seemingly suboptimal biological constraints can lead to efficient, noise-robust decision-making strategies.

Summary updated

BibTeX
@inproceedings{drevet2022efficient,
  author = {Julie Drevet and Jan Drugowitsch and Valentin Wyart},
  title = {{Efficient stabilization of imprecise statistical inference through conditional belief updating}},
  booktitle = {Nature Human Behaviour},
  year = {2022},
  doi = {10.1038/s41562-022-01445-0},
}
Figure