Dynamical information processing in a neuronal microcircuit

我们的大脑由被称为神经元的大量细胞相互连接形成的电路组成。在大脑皮层，占主导地位的神经元类型是锥体细胞，它是我们神经元网络中的主要信息处理器。锥体细胞被数量较少但更多样化的中间神经元所包围，这些中间神经元与锥体细胞及其自身在局部形成联系。通过建立详细的大脑回路的计算机模型，我们将探索这些中间神经元的微回路如何控制通过锥体细胞的信息流。特别是，我们将研究中间神经元如何影响锥体细胞是基于先前存储的知识处理信息，还是从新的经验中学习。知识储存在神经元之间的连接强度中，学习取决于这些突触(连接)强度的可塑性或可变性。人工神经网络(ANN)领域已经证明，可以使用这种范例来构建信息处理设备。然而，目前的神经网络模型使用的电路和细胞类型比我们的大脑简单得多。我们希望进一步了解大脑皮层中发现的复杂微电路的运作，以及它是如何动态控制信息处理和学习的。根据我们的发现，神经网络的新设计应该会出现，它们在处理现实世界的信息处理方面更加灵活和强大。我们将尝试模拟大脑的一小部分如何充当智能记忆设备。我们不断地受到感官信息的轰炸，其中一些我们记得，一些激发了对旧记忆的回忆。大脑如何存储和回忆信息的某些方面被称为联想记忆的数学神经网络模型捕捉到，该模型最早是在40多年前开发出来的。这些模型的工作原理是，通过改变简单计算单元之间的连接强度来存储信息模式，简单计算单元以非常简单的方式模拟大脑中神经元的操作。当先前存储的图案的嘈杂或部分版本被呈现给网络时，旧记忆被调用。这些设备不太灵活。他们必须被告知何时存储模式，以及何时应该回忆起一段记忆。他们可以存储的信息类型相当有限。我们的目标是建立一个更灵活的模型，它可以自己控制以不可预测的速度到达的信息模式的存储和回忆。这个数学模型将基于我们现在所知的海马体神经元电路的许多细节，海马体是哺乳动物大脑的一部分，起着短期记忆的作用。该模型将在计算机软件中实现，并通过在内存中运行存储和回忆的模拟来进行测试。通过建立这个模型，我们希望对许多不同类型的神经元及其形成的复杂电路的实际工作方式有基本的了解。在整个大脑皮层都发现了非常相似的神经元类型和回路，所以我们所学到的应该会增加我们对整个大脑中信息处理的理解，而不仅仅是海马体中的记忆形成。该模型应该为理解和治疗涉及记忆障碍的神经退行性疾病提供相关的见解，例如阿尔茨海默病和各种形式的痴呆症。这项工作也应该对移动机器人领域有价值，因为移动机器人的目标是建立必须与动态环境互动的自主移动机器，就像动物一样。这可能会导致基于神经网络的动态记忆模型的形成，适合纳入到这样的机器人中。

Our brains consist of electrical circuits formed by the interconnection of vast numbers of cells called neurons. In the cortex the dominant neuronal type is the pyramidal cell, which is the main information processor in our neuronal networks. The pyramidal cells are surrounded by a smaller number but much more diverse population of interneurons that form connections locally with pyramidal cells and themselves. By building detailed computer models of brain circuits we will explore how these microcircuits of interneurons control the flow of information through pyramidal cells. In particular we will investigate how the interneurons can influence whether the pyramidal cells are processing information on the basis of previously stored knowledge or are learning from new experiences. Knowledge is stored in the strengths of connections between neurons and learning depends on how plastic or changeable these synaptic (connection) strengths are. The field of artificial neural networks (ANNs) has demonstrated that information processing devices can be built using this paradigm. However, current ANN models use much simpler circuitry and cell types than our brains. We hope to further our understanding of the operation of the complex microcircuitry found in cortex and how it acts to dynamically control information processing and learning. Based on what we discover new designs for ANNs should emerge that are much more flexible and robust in being able to cope with real-world information processing.We will attempt to model how a small section of the brain can act as an intelligent memory device. We are continually bombarded with sensory information, some of which we remember and some of which sparks the recall of old memories. Some aspects of how the brain may store and recall information are captured in mathematical ANN models known as associative memories, which were first developed over 40 years ago. These models work by storing patterns of information via changes in the strengths of connections between simple computing units that mimic the operation of neurons in the brain in a very simple way. Old memories are recalled when a noisy or partial version of a previously stored pattern is presented to the network. These devices are not very flexible. They must be told when to store a pattern and when they are supposed to recall a memory. The type of information they can store is quite limited. We aim to build a much more flexible model that can control for itself the storage and recall of patterns of information arriving at unpredictable rates. This mathematical model will be based upon the many details we now know of the neuronal circuitry of the hippocampus, a part of the mammalian brain that acts as a short-term memory. The model will be implemented in computer software and tested by running simulations of storage and recall in the memory.By building this model we hope to gain fundamental insights into how the many different types of neurons, and the complex circuits they form, actually work. Very similar neuronal types and circuits are found throughout the cortex, so what we learn should increase our understanding of information processing throughout the brain, not just for memory formation in the hippocampus. The model should provide insights of relevance to the understanding of, and therapies for neurodegenerative diseases that involve memory impairment, such as Alzheimer's disease and various forms of dementia.The work should also be of value to the field of mobile robotics where the aim is to build autonomous, mobile machines that must interact with a dynamic environment, in the same way that animals do. It could lead to the formulation of neural network-based dynamic memory models suitable for incorporation into such robots.