Predictive-Coding Neural Networks: How They Learn and Behave

大脑是一个令人困惑的复杂器官，但在解开它的秘密方面正在取得进展。预测编码(PC)理论指出，大脑是一个层级结构，其中每一层只与下面的层和上面的层进行交流。这种结构模仿了哺乳动物大脑皮层中大脑区域之间的相互连接。PC网络中的信息在网络中同时上行和下行。我们的大脑也是这样做的，这表明知觉是一个主动的、双向的推理过程。我们假设，这种双向过程会使感知更加稳健。PC网络仍面临许多挑战。首先，PC分类网络的一些有益特性被认为源于它们的生成性。不是将“2”的图像馈送到网络的底部，并让网络将其分类为2，而是可以在网络的顶部馈入“2”的类向量，并让网络生成2的图像。不幸的是，这些图像通常与训练数据不相似，但看起来是随机的。我们将研究各种技术来鼓励PC网络生成可识别的输入。生成性PC网络在感知方面可能比人工神经网络(ANN)更健壮。当前的ANN很容易被愚弄，只需对输入进行轻微调整。然而，这些敌对的输入并不能愚弄人类。我们将针对敌意输入测试我们的生成性PC网络，看看它们是否会成为相同攻击的受害者，或者它们是否更接近人类的感知。最近发现，PC网络类似于反向传播学习算法；当它们运行时，它们自主学习，不像ANN需要外部管理过程来管理学习。这表明，大脑中可能正在发生某种类似于后备支持的事情。然而，到Backprop的链接被证明只适用于受限的分层体系结构。我们将扩展这一理论，为其他常见的网络拓扑(例如具有循环连接的层)推导出生物学上合理的学习算法。这样的一层可以模拟吸引器网络并实现动态的感知过程，包括类似SoftMax的观看光学错觉的双稳态过程。在这个研究项目中开发的学习算法将给人工智能带来革命性的变化，并为认知神经科学提供信息。这些强大的神经网络将使关键任务应用(如无人驾驶汽车的视觉系统)更加可靠。此外，这些学习方法是自主的，因此PC网络可以在没有监督程序的情况下进行学习，因此更容易部署在专门的硬件中，比如正在新兴的神经形态芯片行业开发的那些硬件。参与这项研究的学生将在人工智能和神经科学领域拥有战略优势，并将成为产业革命的领导者，扩大学术研究的理论和实践。

The brain is a bewilderingly complex organ, but progress is being made to unlock its secrets. The theory of Predictive Coding (PC) states that the brain is a hierarchy of layers in which each layer communicates only with the layer below and the layer above. This architecture mimics the reciprocal connectivity found between brain regions in the mammalian cortex. Information in PC networks flows both up and down the network simultaneously. Our brains do the same, suggesting that perception is an active, two-way inference process. We hypothesize that this two-way process makes perception more robust. A number of challenges remain for PC networks. Firstly, some of the beneficial properties of PC classification networks are thought to stem from them being generative. Instead of feeding an image of a "2" into the bottom of the network and having the network classify it as a two, one can also feed in the class vector for "two" at the top of the network and have the network generate an image of a two. Unfortunately, those images typically do not resemble the training data, but appear random. We will investigate a variety of techniques to encourage PC networks to generate recognizable inputs. Generative PC networks might be more robust at perception than artificial neural networks (ANNs). Current ANNs are easy to fool by adding slight adjustments to the input. Yet, these adversarial inputs do not fool humans. We will test our generative PC networks against adversarial inputs to see if they fall victim to the same attacks, or if they are closer to human perception. It was recently revealed that PC networks approximate the backpropagation learning algorithm; as they run, they learn autonomously, unlike ANNs that require an external governing process to manage the learning. This suggests that something akin to backprop could be going on in the brain. However, the link to backprop was proven only for a restricted, layered architecture. We will extend the theory to derive biologically plausible learning algorithms for other common network topologies, such as layers with recurrent connections. Such a layer could model an attractor network and implement the dynamic process of perception, including the SoftMax-like bistable process of looking at an optical illusion. The learning algorithms developed in this research program will revolutionize AI and inform cognitive neuroscience. These robust neural networks will make mission-critical applications (like vision systems for driverless cars) more reliable. Moreover, these learning methods are autonomous, so PC networks can learn without a supervising program and, as a result, would be easier to deploy in specialized hardware, like those being developed in the emerging industry of neuromorphic chips. The students involved in this research will have a strategic advantage in the fields of AI and neuroscience, and will be leaders in the industrial revolution and expand the theory and practice in academic research.