Neural signatures of perceptual inference.
Neural signatures of perceptual inference.
复制标题
DOI：
10.7554/elife.11476
复制
发表时间：
2016-03-07
期刊：
影响因子：
7.7
通讯作者：
Griffiths TD
中科院分区：
文献类型：
--
作者：
Sedley W;Gander PE;Kumar S;Kovach CK;Oya H;Kawasaki H;Howard MA;Griffiths TD
生成模型，如预测编码，认为感知是由感官输入和先前预测的组合产生的，每一个都由其精度（逆方差）加权，这些称为预测误差（预测偏差）或惊喜（感官输入的负对数概率）之间的不一致性。然而，这种系统的直接证据，以及其计算的生理基础，是缺乏的。使用音高值根据特定规则变化的听觉刺激，我们控制并分离了感知的三个关键计算变量，并使用人类听觉皮层的直接记录发现，由于预测违规而引起的惊讶是由伽马波段（>30 Hz）的局部场电位振荡编码的，β波段的预测变化（12-30 Hz），预测的精度似乎与α波段振荡（8-12 Hz）定量相关。这些结果证实振荡代码的生成模型的感知的关键方面。DOI：http://dx.doi.org/10.7554/eLife.11476.001我们对世界的感知不仅仅基于我们感官的输入。相反，我们所感知的东西也会受到感知的环境和我们对它的期望的严重影响，一些研究人员认为，感知是将我们的感官和预测信息结合起来的结果。这一学派被称为“预测编码”，基本上认为大脑存储了一个世界模型，并将其与来自我们感官的信息进行权衡，以确定我们所感知的。然而，大脑以这种方式工作的直接证据仍然缺失。虽然神经科学家已经看到，当预期和传入的感官信息不匹配时，大脑会做出反应，但没有人在大脑中观察到预测本身。Sedley等人现在通过直接记录接受癫痫手术的志愿者大脑中的电活动，为预测即将到来的感觉信息提供了直接证据。这项实验使用了一种新的方法，志愿者们听了一系列半可预测的声音。也就是说，一开始，志愿者们听到了一组音调相似的声音。在随机间隔之后，这些声音的平均音高发生了变化，在再次随机变化之前，它们变得或多或少可变。这种方法意味着志愿者必须在整个实验过程中不断更新他们的预测。与以前的研究一致，意外的声音，导致感官信息和大脑预测之间的不匹配，与高频脑电波有关。然而，Sedley等人发现，更新预测本身与中频脑电波有关;这证实了预测编码模型的建议。最后，这种方法还意外地揭示了志愿者对预测的信心与低频脑电波有关。在未来，这种新方法将提供一种简单的方法来直接研究人类的感知元素，而且由于实验不需要复杂的学习，其他动物也可以。DOI：http://dx.doi.org/10.7554/eLife.11476.002网站
Generative models, such as predictive coding, posit that perception results from a combination of sensory input and prior prediction, each weighted by its precision (inverse variance), with incongruence between these termed prediction error (deviation from prediction) or surprise (negative log probability of the sensory input). However, direct evidence for such a system, and the physiological basis of its computations, is lacking. Using an auditory stimulus whose pitch value changed according to specific rules, we controlled and separated the three key computational variables underlying perception, and discovered, using direct recordings from human auditory cortex, that surprise due to prediction violations is encoded by local field potential oscillations in the gamma band (>30 Hz), changes to predictions in the beta band (12-30 Hz), and that the precision of predictions appears to quantitatively relate to alpha band oscillations (8-12 Hz). These results confirm oscillatory codes for critical aspects of generative models of perception. DOI: http://dx.doi.org/10.7554/eLife.11476.001 Our perception of the world is not only based on input from our senses. Instead, what we perceive is also heavily altered by the context of what is being sensed and our expectations about it. Some researchers have suggested that perception results from combining information from our senses and our predictions. This school of thought, referred to as “predictive coding”, essentially proposed that the brain stores a model of the world and weighs it up against information from our senses in order to determine what we perceive. Nevertheless, direct evidence for the brain working in this way was still missing. While neuroscientists had seen the brain respond when there was a mismatch between an expectation and incoming sensory information, no one has observed the predictions themselves within the brain. Sedley et al. now provide such direct evidence for predictions about upcoming sensory information, by directly recording the electrical activity in the brains of human volunteers who were undergoing surgery for epilepsy. The experiment made use of a new method in which the volunteers listened to a sequence of sounds that was semi-predictable. That is to say that, at first, the volunteers heard a selection of similarly pitched sounds. After random intervals, the average pitch of these sounds changed and they became more or less variable for a while before randomly changing again. This approach meant that the volunteers had to continually update their predictions throughout the experiment In keeping with previous studies, the unexpected sounds, which caused a mismatch between the sensory information and the brain’s prediction, were linked to high-frequency brainwaves. However, Sedley et al. discovered that updating the predictions themselves was linked to middle-frequency brainwaves; this confirms what the predictive coding model had suggested. Finally, this approach also unexpectedly revealed that how confident the volunteer was about the prediction was linked to low-frequency brainwaves. In the future, this new method will provide an easy way of directly studying elements of perception in humans and, since the experiments do not require complex learning, in other animals too. DOI: http://dx.doi.org/10.7554/eLife.11476.002