本文为美国加州理工学院（作者：Timothy X Brown）的博士论文，共109页。

神经网络是一组高度互联的简单处理器，这些许多的连接允许信息在网络中快速传输，由于其简单性，在一个网络中构建许多处理器是可行的。这些特性意味着我们可以利用神经网络建立高效的大规模并行机器，主要问题是我们如何在神经网络中进行互连。迄今为止开发的各种方法（如外积、学习算法或能量函数）都存在以下缺陷：训练/规范时间长；不能保证在所有输入上工作；需要完全连接。

我们讨论了使用问题本身的拓扑和约束来设计神经解的拓扑和连接方法。我们定义了几个有用的电路（Winner-Take-All电路），这允许我们以一种受控的方式使用反馈来引入约束。这些电路被证明是稳定的，并且只收敛于有效状态。我们使用Hopfield电子模型，因为这接近于实际的实现。我们还讨论了将这些电路合并到更大系统中的方法，包括神经和非神经的方式。通过利用我们定义中的规律，可以构建有效的网络。

为了演示这些方法，我们从通信中寻求三个问题。本文首先讨论了电路切换问题的两种应用：大型多级开关的寻路问题和呼叫重排问题。这两种应用都说明了我们如何利用许多神经元来构造大规模的并行机器，以及Winner-Take-All如何能够简化我们的设计。接下来，我们开发了一种解决高速分组交换机仲裁问题的方案。定义了一类有用的交换网络，然后设计了一个神经网络来解决这类问题的竞争仲裁问题。分析了神经网络/开关系统的各个方面，测量了该方法的排队性能。利用基本设计提出了一种适用于大型（1024输入）ATM分组交换机的可行架构。利用神经网络的大规模并行性，我们可以考虑以前无法计算的算法，这些现在可行的算法使我们对交换机设计有了新的想法。

A neural network is a highly interconnectedset of simple processors. The many connections allow information to travelrapidly through the network, and due to their simplicity, many processors inone network are feasible. Together these properties imply that we can buildefficient massively parallel machines using neural networks. The primaryproblem is how do we specify the interconnections in a neural network. Thevarious approaches developed so far such as outer product, learning algorithm,or energy function suffer from the following deficiencies: longtraining/specification times; not guaranteed to work on all inputs; requiresfull connectivity.

Alternatively we discuss methods of usingthe topology and constraints of the problems themselves to design the topologyand connections of the neural solution. We define several usefulcircuits—generalizations of the Winner-Take-All circuit— that allows us toincorporate constraints using feedback in a controlled manner. These circuitsare proven to be stable, and to only converge on valid states. We use theHopfield electronic model since this is close to an actual implementation. Wealso discuss methods for incorporating these circuits into larger systems,neural and nonneural. By exploiting regularities in our definition, we canconstruct efficient networks.

To demonstrate the methods, we look tothree problems from communications. We first discuss two applications toproblems from circuit switching; finding routes in large multistage switches,and the call rearrangement problem. These show both, how we can use manyneurons to build massively parallel machines, and how the Winner-Take-Allcircuits can simplify our designs. Next we develop a solution to the contentionarbitration problem of high-speed packet switches. We define a useful class ofswitching networks and then design a neural network to solve the contentionarbitration problem for this class. Various aspects of the neural network/switchsystem are analyzed to measure the queueing performance of this method. Usingthe basic design, a feasible architecture for a large (1024-input) ATM packetswitch is presented. Using the massive parallelism of neural networks, we canconsider algorithms that were previously computationally v unattainable. Thesenow viable algorithms lead us to new perspectives on switch design.

1. 引言

2. 神经网络设计

3. 电路交换网络的控制

4. Banyan网络控制器

5. 后记

更多精彩文章请关注公众号：