基于高效编码假说和贝叶斯推断理论构建的计算模型已成为视觉认知研究领域的重要理论框架。该模型不仅能够有效解释视知觉的多种行为表现, 更重要的是, 其提供了一个统一的计算框架来系统揭示视觉信息处理完整过程。本文系统综述了该模型的理论基础及其在视觉认知研究中的应用进展。基于现有研究, 未来可结合神经电生理学技术进一步验证该模型的神经生理基础, 并拓展其在其他感官知觉, 高级社会认知和特殊人群等领域的研究。

The efficient-coding constrained Bayesian inference model, developed by Stocker and colleagues (Wei & Stocker, 2012, 2015, 2017; Sun, Xu & Stocker, 2025), has emerged as a pivotal theoretical framework in visual cognition research. This model delineates a comprehensive perceptual process comprising three key stages: efficient sensory encoding, Bayesian decoding, and perception-action mapping. It formally characterizes the interplay between neural resource allocation (Fisher Information, $ J(\theta)$), prior ($P(\theta)$), width of neural tuning curves/likelihood distribution $\left(\sigma^{2}\right)$, perception biases ($b(\theta)$), as expressed by the following relationships:$ p(\theta) \propto \sqrt{J(\theta)} ; \sigma^{2} \geq 1 / \sqrt{J(\theta)} \propto 1 / p(\theta) ; b(\theta) \propto\left(\frac{1}{p(\theta)^{2}}\right)^{\prime}$

Collectively, these equations unveil the computational principles governing visual perception under environmental constraints.

Existing research has explored the model’s implications from three primary perspectives:

1. Unified Account of Perceptual Biases: The model provides a parsimonious computational framework explaining diverse phenomena, including noise-dependent biases (Wei & Stocker, 2012, 2015), trade-offs between prior efficiency and encoding fidelity (Hahn & Wei, 2024), and response-range effects (Sun et al., 2025). Notably, it maintains predictive consistency across both low-level (e.g., orientation) and high-level (e.g., optic flow) visual features.

2. Multi-Timescale Prior Integration: The framework reveals how long-term priors (reflecting learned statistical regularities) stabilize perception, while short-term priors (driven by recent stimuli) induce serial dependence (Fischer & Whitney, 2014). Dynamic weighting between these mechanisms enables adaptive optimization (Fritsche et al., 2020; Sun et al., 2025).

3. Computational Basis of ASD Perceptual Anomalies: Compared to neurotypical individuals, populations with autism spectrum disorder (ASD) exhibit reduced encoding efficiency and impaired statistical adaptation (Noel et al., 2021; Sapey-Triomphe et al., 2023), offering a novel computational lens for understanding sensory processing differences.

Building on these advances, four critical directions can be considered in future studies:

1. Multisensory and Cross-Modal Applications: Current research indicates that the efficient coding principle is universally observed across unimodal sensory systems, such as vision and audition, while Bayesian decoding mechanisms are also applicable to cross-modal integration (e.g., Ernst & Banks, 2002). However, existing studies integrating these two frameworks remain largely confined to the visual domain. Future work should validate the applicability of the "efficient coding-constrained Bayesian model with perceptual-response mapping modulation" in other sensory modalities and cross-modal contexts, thereby establishing a unified computational framework for perception (Sun et al., 2025).

2. Extensions to Social Cognition: Social cognitive processes are similarly influenced by prior experience and exhibit Bayesian inference properties (e.g., the Prevalence-Induced Concept Change [PICC] phenomenon, Levari et al., 2018). While prior research has confirmed the ubiquity of Bayesian inference in social cognition—including intention understanding, emotion recognition (Baker et al., 2009), and cultural transmission—the role of resource-constrained encoding mechanisms remains unexplored. Future studies should investigate whether social cognitive priors adhere to efficient coding principles, how limited cognitive resources modulate information integration, and how these processes differ from general perceptual mechanisms. Such efforts will advance computational theories of the "social brain."

3. Computational Specificity in Special Populations: Individuals with ADHD, dyslexia, and other neurodevelopmental conditions exhibit atypical perceptual processing, including impaired encoding efficiency or aberrant prior formation (e.g., Nigg & Casey, 2005). Future research should employ psychophysical experiments and computational modeling (e.g., Cramér-Rao lower bound) to quantify these population-specific encoding differences. These findings could inform targeted perceptual training protocols, providing computational biomarkers for clinical diagnosis and theoretical foundations for intervention strategies.

4. Elucidating Neural Mechanisms: Although existing studies demonstrate that neural activity aligns with both efficient coding (Ganguli & Simoncelli, 2014) and Bayesian decoding (Rao, 2004) principles—and fMRI research has preliminarily linked neural activity to behavioral performance (Zhang et al., 2025)—three critical gaps remain: (1) the neural basis of multimodal perception, (2) the synergistic mechanisms between efficient coding and Bayesian inference, and (3) the development of an integrated "behavior-neural-computation" mapping model. Addressing these challenges will provide theoretical foundations for brain-inspired artificial intelligence systems.

Addressing these questions will not only advance theoretical paradigms in perceptual computation but also deepen our understanding of human perception. Moreover, these insights may inspire biologically grounded designs for brain-like artificial intelligence systems, with transformative implications for clinical practice and machine perception.