Common computational principles underlying biases in visual perception, memory and confidence

Abstract

This thesis investigates the nature, occurrence, and underlying mechanisms of biases in visually guided behavior. By studying human visual perception, memory and confidence, I reveal that these aspects of vision are systematically biased in a variety of ways. These biases could be comprehended through normative theories of visual processing—the efficient coding hypothesis and the Bayesian brain hypothesis. Using a theoretical framework about neural representation and interpretation of sensory information, I explain how these biases emerge from optimized encoding and decoding processes tailored to the statistical structures of the surrounding environment. The thesis is structured around three comprehensive studies that explore how the brain leverages prior beliefs to optimize behavior. Chapter 2 examines how memory of visual stimuli is altered by both preceding and following stimuli. I provide empirical evidence demonstrating that they not only bias memory reports but also enhance memory precision, thereby supporting the concept of efficient coding in visual working memory. Chapter 3 focuses on perceptual estimation, uncovering that perceptions of current stimuli are concurrently attracted to and repelled from recent sensory history. A normative computational model developed in this chapter explains that these opposite biases arise from efficient encoding and Bayesian decoding processes. Chapter 4 extends the exploration to higher-level metacognition. I show that visual confidence is systematically biased, reflecting behavioral variability but not the biases. This indicates that even metacognition and consciousness may be optimized in a manner similar to low-level visual perception, adhering to Bayesian principles. The findings presented in this thesis challenge the common belief that biases are malfunctions, instead supporting the notion that they reflect optimal neural information processing. This thesis underscores the importance of considering biases not as errors, but as reflections of the brain’s adaptive strategies in a statistically structured world. It contributes to the comprehensive understanding of biases across visual perception, memory and confidence, and represents a significant step forward in the ongoing effort to unravel the optimality in human behavior.