Destructured sensory fields, involving homogenous stimulation with little or no time-varying structure, provide a fertile ground for testing hypotheses about predictive coding in the human brain. Extended exposure to sensory patterns that deviate substantially from the statistics of natural environments can elicit a bewildering range of perceptual phenomena, up to and including vivid oneiric imagery. We illustrate how this large variety of perceptual effects can be understood as the experiential counterpart of auto-generated neuronal dynamics, unconstrained by parameters that tune the waking sensorium. We synthesize the literature on autonomous neuronal activity across multiple spatiotemporal scales with generative models of brain function and evidence from artificial neural architectures. Perception, we argue, emerges from a process of non-random sampling from an intrinsic distribution of hypotheses rather than a direct transfer of information from the world. The imagery that occurs in altered sensory environments is explained as the outcome of an iterative search through internal world models in which the structural typology of percepts reflects the brain’s intrinsic functional architectures.

涉及均匀刺激而几乎没有或没有随时间变化的结构的变形的感觉场为测试关于人脑中预测编码的假设提供了沃土。长时间暴露于与自然环境统计数据大相径庭的感觉模式，可能会引起令人困惑的知觉现象，直至并包括生动的自拍影像。我们说明了如何将这种各种各样的知觉效果理解为自动生成的神经元动力学的经验对应物，不受调节苏醒的感觉层的参数的约束。我们通过脑功能的生成模型和来自人工神经结构的证据，综合了跨多个时空尺度的自主神经元活动的文献。我们认为知觉，从假设的内在分布而不是从世界直接传递信息的非随机抽样过程中产生的。发生在改变的感觉环境中的图像被解释为通过内部世界模型进行迭代搜索的结果，在内部世界模型中，感知的结构类型反映了大脑的内在功能结构。