Optical Imaging Group, Institut für Neuroinformatik, Ruhr University Bochum, Bochum, Germany.
Institute for Frontier Areas of Psychology and Mental Health, Freiburg, Germany.
PLoS One. 2020 May 4;15(5):e0232349. doi: 10.1371/journal.pone.0232349. eCollection 2020.
A popular model for sensory processing, known as predictive coding, proposes that incoming signals are iteratively compared with top-down predictions along a hierarchical processing scheme. At each step, error signals arising from differences between actual input and prediction are forwarded and recurrently minimized by updating internal models to finally be "explained away". However, the neuronal mechanisms underlying such computations and their limitations in processing speed are largely unknown. Further, it remains unclear at which step of cortical processing prediction errors are explained away, if at all. In the present study, human subjects briefly viewed the superposition of two orthogonally oriented gratings followed by abrupt removal of one orientation after either 33 or 200 milliseconds. Instead of strictly seeing the remaining orientation, observers report rarely but highly significantly an illusory percept of the arithmetic difference between previous and actual orientations. Previous findings in cats using the identical paradigm suggest that such difference signals are inherited from first steps of visual cortical processing. In light of early modeling accounts of predictive coding, in which visual neurons were interpreted as residual error detectors signaling the difference between actual input and its temporal prediction based on past input, our data may indicate continued access to residual errors. Such strategy permits time-critical perceptual decision making across a spectrum of competing internal signals up to the highest levels of processing. Thus, the occasional appearance of a prediction error-like illusory percept may uncover maintained flexibility at perceptual decision stages when subjects cope with highly dynamic and ambiguous visual stimuli.
一种流行的感觉处理模型,称为预测编码,提出输入信号沿着分层处理方案与自上而下的预测进行迭代比较。在每一步,实际输入和预测之间的差异产生的误差信号通过更新内部模型来传递和递归最小化,最终被“解释掉”。然而,这种计算的神经元机制及其在处理速度方面的局限性在很大程度上是未知的。此外,如果有的话,在皮层处理的哪个步骤预测误差被“解释掉”仍然不清楚。在本研究中,人类受试者短暂地观察了两个正交取向的光栅的叠加,然后在 33 或 200 毫秒后突然去除一个取向。观察者报告说,他们很少但非常显著地看到了前一个取向和实际取向之间算术差的错觉,而不是严格地看到剩余的取向。以前在使用相同范式的猫身上的发现表明,这种差异信号是从视觉皮层处理的第一步继承而来的。鉴于预测编码的早期建模解释,其中视觉神经元被解释为基于过去输入的实际输入与其时间预测之间的差异的残余误差检测器,我们的数据可能表明对残余误差的持续访问。这种策略允许在一系列竞争的内部信号中进行时间关键的感知决策,直到处理的最高水平。因此,当受试者应对高度动态和模糊的视觉刺激时,预测误差样错觉的偶尔出现可能揭示了感知决策阶段的灵活性。