A unifying theory explains seemingly contradicting biases in perceptual estimation

Xue-Xin Wei, Michael Hahn

TL;DR

This paper introduces a unified Bayesian theory that explains seemingly contradictory biases in human perception. By decomposing perceptual biases into attractive and repulsive components influenced by prior expectations, encoding precision, and loss functions, the authors provide a simple rule to predict bias direction and apply it to diverse perceptual tasks, offering new insights into how our brains process sensory information.

What Problem Does This Paper Address?

Human perception is not a perfect, objective reflection of reality; it's often biased. For decades, researchers have identified various types of perceptual biases, such as attraction to prior expectations (e.g., assuming light always comes from above) and repulsion from these priors (e.g., a bias towards oblique lines in vision, despite them being less frequent in natural scenes). These seemingly contradictory findings have made it challenging to develop a cohesive understanding of how perceptual biases arise. This paper aims to bridge this gap by proposing a unifying theoretical framework that can explain these diverse and sometimes conflicting biases within a single, coherent model.

Methods

The core of the paper is a novel Bayesian observer model that allows for arbitrary priors and encoding strategies. The authors analytically derive an "additive decomposition" of perceptual biases into three key components:

  1. Prior Attraction: A tendency to perceive stimuli closer to one's prior expectations.
  2. Likelihood Repulsion: A tendency to be repelled from regions where sensory encoding is highly precise, often linked to efficient coding principles.
  3. Regression towards the Boundary: A tendency to regress towards the center of a bounded stimulus range.

This decomposition leads to a simple "Prior/Precision ratio rule" that predicts the direction of perceptual bias. They also developed a general numerical fitting procedure to estimate the model's components (prior, encoding, and noise) from trial-by-trial behavioral data, allowing them to account for both bias and response variability.

Datasets

The authors applied their model to several existing datasets from previous experiments, covering a range of perceptual tasks:

  • Orientation Estimation: Data from de Gardelle et al. (38) and Tomassini et al. (39) / Girshick et al. (41).
  • Perceptual Learning of Motion Direction: Data from Gekas et al. (67).
  • Magnitude Perception (Numerosity and Time Interval Estimation): Data from Xiang et al. (66) and Remington et al. (63).
  • Color Perception: Data from Bae et al. (65) and Olkkonen et al. (64).

No new datasets were introduced; instead, the paper demonstrates the model's applicability to diverse existing benchmarks.

Key Insights and Ideas

  • Unified Theory of Biases: The most significant insight is the unifying framework that explains seemingly contradictory biases (attraction vs. repulsion) as arising from the interplay of prior beliefs and the precision of sensory encoding.
  • Prior/Precision Ratio Rule: A simple, universal rule emerges for predicting bias direction, based on how quickly the prior and encoding precision change relative to each other across the stimulus space. This is a powerful predictive tool.
  • Encoding Heterogeneity is Crucial: For circular variables (like orientation or motion direction), encoding heterogeneity (i.e., not all parts of the stimulus space are encoded with equal precision) plays a surprisingly dominant role in driving biases, often more so than prior attraction.
  • Loss Function Matters for Circular Variables: The choice of loss function (how errors are penalized) significantly impacts biases for circular stimuli, with higher exponents leading to stronger repulsive biases. This is less critical for scalar variables.
  • Distinction Between Circular and Scalar Variables: The paper highlights a fundamental difference in how biases manifest for circular (e.g., orientation) versus scalar (e.g., magnitude) variables, with encoding heterogeneity being more critical for the former.

Key Results

  • The proposed unifying theory successfully accounts for perceptual biases across a wide range of tasks, including orientation, motion direction, color, numerosity, and time interval estimation.
  • The Prior/Precision ratio rule accurately predicts the sign of observed biases in the analyzed datasets.
  • For circular variables, repulsive biases (driven by encoding heterogeneity) often play a larger role than prior attraction.
  • The model's fitting procedure, which estimates priors, encoding, and noise from trial-by-trial data, provides a more comprehensive account of perceptual behavior than previous models.
  • In orientation perception, the inferred loss function exponent was surprisingly high (p=8), suggesting observers are more tolerant of small errors than previously assumed.

Limitations and Future Work

  • Loss Function Detail: While the paper considers an LP family of loss functions, it acknowledges that future work could explore more detailed, potentially asymmetric, loss functions.
  • Multi-dimensional Stimuli: The current theory is primarily one-dimensional. Extending it to multi-dimensional cases (e.g., biases in 3D space perception) is an important future direction.
  • Adaptation and Context: The model does not fully account for adaptation or contextual effects, which could influence encoding and perception.
  • Dynamic Evolution of Priors and Encoding: The paper raises the question of how priors and encoding dynamically adapt to environmental statistics, suggesting this is an open area for future research.
  • Neural Implementation: While the theory informs neural implementations of Bayesian inference, understanding how specific neural circuits implement priors, loss functions, and rapid prior adjustments remains a challenge.

Summary updated

Abstract

Human perception is often biased. As key signatures of the underlying computational process, these biases are of tremendous interest to brain scientists. Over the past decades, Bayesian inference has emerged as a major theoretical framework for understanding perception, and more generally cognition. To understand these biases, previous work has proposed a number of conceptually different and even seemingly contradicting ingredients, including attraction to a Bayesian prior, repulsion from the prior due to efficient coding, and central tendency effects on a bounded range. We present a unifying Bayesian theory of biases in perceptual estimation. We theoretically demonstrate an additive decomposition of perceptual biases into attraction to a prior, repulsion away from regions with high encoding precision, and regression away from the boundary. Importantly, the results reveal a simple and universal rule for predicting the direction of perceptual biases. To fit our Bayesian framework to the data, we developed a general numerical fitting procedure that estimates the model components, including prior, encoding, and noise magnitude, by maximizing the likelihood of trial-by-trial response data. Hence, the model is fitted to account for the full response distribution conditioned on the stimulus, taking both the perceptual bias and the response variability into account. We applied our modeling framework to a number of datasets collected in previous experiments, including those that investigated color perception, orientation perception, perceptual learning of motion direction, the estimation of numerosity, and the estimation of the length of time intervals. Applications to these experimental datasets reveal both domain-specific and general insights. In particular, the results reveal a major difference between the perceptual biases of circular variables and scale variables. Overall, our theory accounts for and leads to new understandings of biases in the perception of a variety of stimulus attributes.
BibTeX
@article{wei2023unifying,
  author = {Xue-Xin Wei and Michael Hahn},
  title = {{A unifying theory explains seemingly contradicting biases in perceptual estimation}},
  journal = {Journal of Vision},
  year = {2023},
  doi = {10.1167/jov.23.9.5082},
}
Figure