The Dynamics of Top-Down Attention: Predictive Coding Loops for Auditory Selective Attention

1980年，著名的视觉科学家理查德·格雷戈里阐述了一个关于人类知觉的非凡观点。他说，感知并不是我们的感官系统向我们的大脑发送信息的直接结果。相反，感知是对世界的一种心理模型，我们的大脑只是检查我们的感官，以确保该模型不会产生错误。*这种想法背后是贝叶斯推理的概念：感知某事的最佳方式是随着时间的推移积累和整合有关它的统计证据。尽管格雷戈里的提议似乎违反直觉，与我们的经验相反，但它与关于生物和人工系统如何感知世界的新证据非常吻合。最近机器学习领域的巨大进步将这一理念应用于能够感知语音和理解视频的神经网络中，而神经学家正在发现证据表明，人类大脑在一种被称为预测编码的过程中使用了相同的基本计算。*预测编码可能对听觉感知特别重要，因为到达我们耳朵的声音通常是来自许多不同来源的信号的混合。我们想要听到的声音需要是“纯混合的”，而预测编码提供了一种关于大脑如何做到这一点的理论。这项研究采用了三种方法来研究我们即使在有大量干扰的情况下也能选择性地听一种声音的能力：1)开发和测试听觉预测编码神经网络的计算神经科学程序；2)使用神经成像技术了解大脑如何通过预测编码实现选择性注意力的认知神经科学程序；3)认知机器人程序，将这些想法转化为软件，使机器人能够在嘈杂的现实环境中对语音命令做出反应。*这项工作将对加拿大产生直接和长期的好处。通过研究和更好地理解人脑的基本计算，我们可以为机器人和医疗设备等广泛的应用发明更好的技术。我们还可以开始了解科学中最深奥的奥秘之一：大脑的生物学是如何产生思维的感知的？

In 1980, the eminent vision scientist Richard Gregory articulated a remarkable idea about human perception. He said that perceiving is not the direct result of our sensory systems sending information into our brains. Instead, perception is a mental model of the world, and our brains merely check with our senses to make sure the model isn't producing errors.***Underlying this idea is the concept of Bayesian inference: that the best way to perceive something is to accumulate and integrate statistical evidence about it over time. Although Gregory's proposal seems counterintuitive and contrary to our experience, it closely matches emerging evidence about how both biological and artificial systems can perceive the world. Recent dramatic advances in machine learning use this idea in neural networks that can perceive speech and understand video, while neuroscientists are finding evidence that the human brain uses the same kind of fundamental computation in a process called predictive coding.***Predictive coding is probably especially important for auditory perception because the sounds reaching our ears are usually mixtures of signals from many different sources. The sound we want to hear needs to be “unmixed” and predictive coding offers a theory about how the brain might do this. The research proposed here takes three approaches to study our ability to selectively listen to one sound even when there is lots of distraction: 1) a computational neuroscience program to develop and test predictive coding neural networks for hearing; 2) a cognitive neuroscience program using neuroimaging to understand how the brain implements selective attention by predictive coding; and 3) a cognitive robotics program that translates these ideas into software to make robots that can respond to speech commands in noisy real-world environments.***This work will have both immediate and long-term benefits for Canada. By studying and better understanding the fundamental computations of the human brain, we can invent better technologies for a wide range of applications such as robotics and medical devices. We can also begin to understand one of the most profound mysteries of science: how does the biology of the brain generate the percepts of the mind?