Spaun: A Perception-Cognition-Action Model Using Spiking Neurons
Spaun: A Perception-Cognition-Action Model Using Spiking Neurons
复制标题
Spaun：使用尖峰神经元的感知-认知-行动模型
DOI：
--
复制
发表时间：
2012
期刊：
影响因子：
--
通讯作者：
C. Eliasmith
中科院分区：
文献类型：
--
作者：
T. Stewart;Feng;C. Eliasmith
Spaun：一个使用spike神经元的感知-认知-行动模型Terrence C. Stewart (tcstewar@uwaterloo.ca) fengi - xuan Choo (fchoo@uwaterloo.ca) Chris Eliasmith （celiasmith@uwaterloo.ca）滑铁卢大学理论神经科学中心，滑铁卢，安大略省，N2L 3G1摘要我们提出了一个大规模的认知神经模型Spaun（语义指针架构）。，并展示了6个任务（数字识别、记忆溯源、串行工作记忆、答题、计数加法、符号模式补全）的仿真结果。该模型由230万个尖峰神经元组成，其神经特性、组织和连通性与哺乳动物的大脑相匹配。输入包括手写和输入的数字和符号的图像，输出是一个2自由度的手臂的运动，它写模型的响应。任务可以以任何顺序呈现，而无需为每项任务“重新布线”。相反，该模型能够进行内部认知控制（通过基底神经节），有选择地在大脑中传递信息，并根据每个任务的需要招募不同的皮层成分。模型的输入包括理想化的和手写的数字和符号（图1a）。这些图像是一个28x28的像素网格。这个输入域的优点是包含了显著可变的真实输入，同时还提供了模型必须推理的合理有限的语义。该模型的输出包括一个2自由度手臂的运动。神经模型生成目标位置序列，直接驱动手臂控制器。这为模型提供了自己的手写输出（图1b）。关键词：神经工程；认知体系结构;强化神经元;认知控制;在即将出版的一本书中，Eliasmith（2012）详细介绍了一种用于生物认知的神经结构，称为语义指针结构（SPA）。这种架构基于神经工程框架（Eliasmith & Anderson, 2003），使用一组尖峰神经元来形成高维向量的分布式表示，这些向量反过来可以编码类似符号的树结构。神经元组之间的突触连接在这些向量上计算特定的功能，允许在详细的尖峰神经元模型中实现高级认知算法。在本文中，我们概述了语义指针体系结构：统一网络（Spaun）模型，并讨论了它在六种不同任务上的行为。我们证明了这种生物学上合理的尖峰神经元模型具有以下特征：任务灵活性：任务之间的模型没有变化。视觉输入指示下一步要执行的任务。运动计划：模型输出提供了一个简单的2关节手臂的运动计划，给出手写的数字作为响应。视觉记忆：即使输入已被识别和分类为特定的符号，原始图像的细节仍然可以恢复和使用。组合性：多个项目可以被表示并可靠地绑定在一起，允许创建和操作类似符号树的结构。符号归纳：视觉输入中的类似语言的模式可以在几次演示后被发现，并用于指导后续的响应。图1:Spaun的视觉输入和手臂运动输出示例。输入数字来自MNIST数据库，输入符号用于通知模型当前任务的详细信息。该模型通过控制一个2关节臂来产生输出，内部表示的变化产生其输出笔迹的变化。我们可以把蜘蛛想象成只有一只固定的眼睛和一只有两个关节的手臂。实验对象的眼睛并没有移动，而是实验者通过显示不同的输入来改变落在眼睛上的图像，每个输入显示150毫秒，然后是150毫秒的空白背景。要开始一项特定的任务，向Spaun展示字母“a”，后面跟着一个0到7之间的数字。随后的输入由模型在指定任务的上下文中解释并进行相应的处理，从而产生提供Spaun响应的手臂运动。所有的内部处理都是通过尖峰神经元完成的，其神经特性和连通性与哺乳动物的大脑一致。
Spaun: A Perception-Cognition-Action Model Using Spiking Neurons Terrence C. Stewart (tcstewar@uwaterloo.ca) Feng-Xuan Choo (fchoo@uwaterloo.ca) Chris Eliasmith (celiasmith@uwaterloo.ca) Centre for Theoretical Neuroscience, University of Waterloo, Waterloo, ON, N2L 3G1 Abstract We present a large-scale cognitive neural model called Spaun (Semantic Pointer Architecture: Unified Network), and show simulation results on 6 tasks (digit recognition, tracing from memory, serial working memory, question answering, addition by counting, and symbolic pattern completion). The model consists of 2.3 million spiking neurons whose neural properties, organization, and connectivity match that of the mammalian brain. Input consists of images of handwritten and typed numbers and symbols, and output is the motion of a 2 degree-of-freedom arm that writes the model’s responses. Tasks can be presented in any order, with no “rewiring” of the brain for each task. Instead, the model is capable of internal cognitive control (via the basal ganglia), selectively routing information throughout the brain and recruiting different cortical components as needed for each task. Input to the model consists of idealized and hand-written digits and symbols (Figure 1a). These images are given as a 28x28 grid of pixels. This input domain has the advantage of including significantly variable, real-world input, while also providing a reasonably limited semantics that the model must reason about. Output from the model consists of the motion of a 2-degree-of-freedom arm. The neural model generates a sequence of target locations which directly drive the controller for the arm. This provides the model with its own handwriting output (Figure 1b). Keywords: Neural engineering; cognitive architecture; spiking neurons; cognitive control; whole-brain systems Introduction In a forthcoming book, Eliasmith (2012) details a neural architecture for biological cognition called the semantic pointer architecture (SPA). This architecture, based on the Neural Engineering Framework (Eliasmith & Anderson, 2003), uses groups of spiking neurons to form distributed representations of high-dimensional vectors, which can in turn encode symbol-like tree structures. Synaptic connections between groups of neurons compute particular functions on those vectors, allowing high-level cognitive algorithms to be implemented in detailed spiking neuron models. In this paper, we present an overview of the Semantic Pointer Architecture: Unified Network (Spaun) model and discuss its behaviour on six different tasks. We demonstrate that this biologically plausible spiking neuron model has the following features: Task Flexibility: No changes are made to the model between tasks. Visual input indicates which task to do next. Motor Plans: Model output provides a motor plan for a simple 2-joint arm, giving hand-written digits as responses. Visual Memory: Even after an input has been recognized and classified as a particular symbol, details of the original image can still be recovered and used. Compositionality: Multiple items can be represented and reliably bound together, allowing for the creation and manipulation of symbol-tree-like structures. Symbolic Induction: Language-like patterns in visual input can be discovered after only a few presentations, and used to guide subsequent responses. Figure 1: Example visual input and arm-movement output from Spaun. Input digits are from the MNIST database, and input symbols are used to inform the model of details of the current task. The model produces output by controlling a 2- joint arm, and variation in the internal representations produces the variation in its output hand-writing. We can think of Spaun as having a single, fixed eye and a single 2-joint arm. The eye does not move, but instead the experimenter changes the image falling on it by showing different inputs over time, with each input shown for 150ms, followed by 150ms of blank background. To begin a specific task, Spaun is shown the letter “A” followed by a number between zero and seven. The subsequent input is then interpreted by the model in the context of the specified task and processed accordingly, resulting in arm movements that provide Spaun’s response. All internal processing is performed using spiking neurons, with neural properties and connectivity consistent with the mammalian brain.