Effectiveness of a Kalman filter model constrained by efficient coding in explaining the coexistence of repulsive and attractive perceptual biases

Human perception, despite its high precision, consistently exhibits systematic biases. Interestingly, perceptual biases measured from perceptual estimation tasks can be either towards (attractive) or away from (repulsive) a feature reference, depending on specific visual features and experimental design. Extensive studies have been conducted to investigate computational mechanisms underlying these biases, many of which have introduced a Bayesian framework to integrate the influence of current input with prior knowledge. This framework is practically successful, especially with constraints of efficient coding of the current visual input. However, most models developed under this framework are mainly data-driven, requiring a large number of free parameters to flexibly fit empirical patterns. This approach, while enabling accurate fitting, often compromises model interpretability. In the present study, we aim to establish a unified model that reconciles repulsive and attractive biases using only two naturally presumed processes—efficient coding and Kalman filtering—along with a minimal set of free parameters. We tested the model by fitting data from three perceptual estimation studies exhibiting distinct bias patterns. The results showed that both attractive and repulsive biases were well captured, supporting the model’s validity and adaptability. Together, these findings suggest that the integration of efficient encoding and recursive Bayesian inference through Kalman filtering provides a parsimonious yet powerful account of a broad spectrum of perceptual biases. The model offers insight into a mechanistic explanation for diverse empirical patterns across perceptual domains.