How do biological neural networks learn to predict their environment?

关于大脑如何工作的最流行的理论认为，动物学习对其环境的预测模型。此外，这样的预测模型使他们能够有效地导航并选择适当的行动来实现他们的目标。神经科学界强烈认为，大脑中发现的高度复杂的神经元网络负责编码这样一个预测模型。然而，目前还不清楚这个网络中的连接结构是如何出现的。在出生时，这个网络严重不发达，无法预测太多，但随着动物探索更多的环境，这种网络的连接结构发生变化，动物开始能够预测并适当地对环境采取行动。这种适应背后的物理机制仍然知之甚少。在这个项目中，我们试图通过模拟生物神经网络的模型并仔细观察它们如何在简单的合成环境中学习预测事件来更好地理解这种适应性机制。学习环境预测模型的主题不仅与神经科学相关，而且对于任何试图对某些现象进行建模的科学学科也是如此。尤其是机器学习，多年来一直在追求这样的想法，并产生了关于预测建模的广泛的数学形式主义。大多数有用的实用模型都是手工制作的，但机器学习的一个主要研究目标是产生一种单一的算法，可以从一些数据中输出任意模型。事实上，上述算法被认为是在大脑中运行，并改变其连接，以匹配环境的真实预测模型。目前机器学习中最流行的算法是在松散地模仿大脑的神经元网络中运行，也被称为人工神经网络(ANN)。然而，这些模型与神经生理学和神经解剖学的观察结果不符，与大脑中运行的生物神经网络非常接近。此外，神经网络不能很好地处理时间相关性，特别是视频感知和预测，也无法解释大脑如何很好地处理它们。这也在机器学习领域提出了一个基本问题，解决方案可能会在机器人等领域产生进步，因为现在很难制造出能够学习环境的预测性和随意性(通过动作进行预测)结构的机器人。在这项研究中，我们将专注于理解大脑，并建立可以在机器学习框架下进行评估的预测系统。我们在简单的合成环境中模拟生物上可信的网络，并试图理解它们如何学习这些简单环境的时空因素，以及它的因果动态。我们的目标不是采取基于目标函数改变连通性结构的优化方法，而是要了解简单的生物学上看似合理的规则如何扰乱网络的连通性，以及这种扰动如何帮助正确的学习。遵循计算神经科学的原理，我们将在线性代数和概率论的框架下对所研究的模型进行分析，以提出关于一般生物神经网络的物理行为的可理解的假设。

The most popular theory of how the brain works states that animals learn a predictive model of their environment. Amongst other things, such predictive model allows them to efficiently navigate and to choose the appropriate actions to reach their goals. The neuroscience community strongly believes that the highly complex network of neurons found in the brain is responsible for encoding such a predictive model. However, it is not yet clear how the connectivity structure in this network emerges. At birth, this network is severely underdeveloped and cannot predict much, but as the animal explores more of it's environment, the connectivity structure of such network changes and the animal starts being able to predict and appropriately act upon the environment. The physical mechanisms underlying this adaptation remain poorly understood. In this project, we try to gain a better understanding of such adaptability mechanisms by simulating models of biological neural networks and carefully observing how they learn to predict events in simple synthetic environments.The topic of learning predictive models of the environment is not only relevant to neuroscience but also for any scientific discipline trying to model some phenomena. Machine learning, in particular, has been pursuing such ideas for many years and has produced an extensive mathematical formalism about predictive modeling. Most useful practical models are crafted by hand but a main research goal of machine learning is to produce a single algorithm that can output an arbitrary model from some data. In fact, a said algorithm is believed to be operating in the brain and changing its connections to match the true predictive model of the environment. The most popular algorithm in machine learning right now operates in networks of neurons that are loosely modeled after the brain, also called Artificial Neural Networks (ANNs). Such models, however, do not match observations made by neurophysiology and neuroanatomy and are a very poor approximation of the biological neural networks operating in the brain. Furthermore, ANNs cannot cope very well with temporal correlations, especially for video perception and prediction, and cannot explain how the brain copes so well with them. This also poses a fundamental issue in the field of machine learning and a solution could produce advances in fields like robotics, since it is very hard nowadays to make robots that can learn the predictive and casual (predict with actions) structure of the environment.In this research, we will focus on understanding the brain but also on building predictive systems which can be evaluated under a machine learning framework. We simulate biologically plausible networks in simple synthetic environments and try to understand how they learn the spatio-temporal factors of those simple environments, as well as its causal dynamics. Instead of taking an optimization approach, which changes the connectivity structure based on an objective function, we aim to understand how simple biologically plausible rules perturb the connectivity of the network and how such perturbations aid correct learning. Following with the principles of computational neuroscience we will analyse the models studied under the frameworks of linear algebra and probability theory to come up with understandable hypothesis about the physical behaviour of general biological neural networks.