A Neural Network Model for Reaction Times

任何类型的计算机都需要时间来生成一个问题的答案。心理操作也需要时间来产生答案。一个多世纪以来，心理学家一直在研究反应时间数据，试图从观察到的反应时间模式中推断出正在进行的心理计算的细节。在过去的十年里，人们对“类脑计算”的兴趣重新抬头，这种计算试图在人工系统中获取神经系统的一些计算能力。这种“神经网络”或“连接主义”模型可能非常强大，速度也非常快，尽管目前还不清楚它们最适合解决什么问题。该项目将研究神经网络在执行简单计算所需时间方面的行为，主要目标是了解人类大脑的运作，将神经网络模拟与人类实验数据在相同任务上进行比较。不同的神经网络架构在执行特定操作所需的时间上差异很大。例如，并行计算机通常表现出与更传统的冯·诺伊曼机器完全不同的反应时间模式，因为它们可以同时执行许多操作。例如，最流行的一类神经网络，多层前馈网络，使用完全并行的操作同步操作。所有的计算基本上都花费相同的时间，尽管有可能在这些模型中使用额外的假设建立时间依赖性。然而，另一类并行神经网络模型，非线性动力系统模型，表现出很强的内在时间依赖性。这种神经网络模型的例子有Hopfield网络、ART模型和本研究中使用的模型——BSB模型。人们对神经网络最感兴趣的是它的计算能力，也就是说，它能得到有趣的问题答案。但理想情况下，人类心理操作的模型应该做两件事，计算答案并花费与人类相同的相对时间。对一个简单任务使用BSB模型的初步模拟表明，在一个研究得很充分的简单问题上，判断两个刺激是相同的还是不同的，并做出适当的反应，BSB模型与实验反应时间模式有合理的匹配。对人类认知的神经网络模型产生的反应时间模式的研究将从简单的任务开始，然后发展到更复杂的问题。一个例子是将该模型应用于求解简单算术问题的答案所需的反应时间。

Computers of any kind take time to produce an answer to a problem. Mental operations also take time to produce answers. Psychologists have studied reaction time data for over a century in an effort to infer from the observed patterns of reaction time the details of the mental computation being performed. The past decade has seen a resurgence of interest in "brain-like computation" which attempts to capture in artificial systems some of the computational power of the nervous system. Such "neural network" or "connectionist" models are potentially very powerful and very fast, though it is not yet clear for what problems they are best suited. This project will study the behavior of neural networks with respect to the time they take to perform simple computations, with the primary objective of understanding the operations of the human brain, comparing neural network simulations to human experimental data on the same tasks. Different neural network architectures differ a great deal in the time it takes for a particular operation to be performed. Parallel computers, for instance, generally exhibit reaction time patterns completely different from more traditional von Neumann machines, because they can perform many operations at the same time. For example, the most popular class of neural networks, multilayer feed forward networks, use completely parallel operations operating in synchrony. All computations take essentially the same time, although it is possible to build time dependencies into these models with additional assumptions. However, another class of parallel neural network models, non- linear dynamical system models, show very strong intrinsic time dependencies. Examples of such neural network models are Hopfield networks, the ART models, and the model to be used in this research, the BSB model. Most of the interest in neural networks has been in their computational abilities, i.e., getting interesting answers to problems. But ideally, a model of a human mental operation should do two things, compute the answer and take the same relative time to do it as humans. Preliminary simulations using the BSB model for a simple task have shown a reasonable match to experimental reaction time patterns in the well-studied, simple problem of deciding whether two stimuli are the same or different and responding appropriately. Study of the reaction time patterns produced by neural network models for human cognition will begin with simple tasks and then progress to more complex problems. One example is application of this model to the reaction time required to obtain the answers to problems in simple arithmetic.