The Proper Treatment of Symbols in a Connectionist Architecture
The Proper Treatment of Symbols in a Connectionist Architecture
复制标题
联结主义架构中符号的正确处理
DOI：
--
复制
发表时间：
2000
期刊：
影响因子：
--
通讯作者：
J. Hummel
中科院分区：
文献类型：
--
作者：
K. Holyoak;J. Hummel
现代认知科学的一个基本原则是物理符号系统假说，它简单地说，人类认知是物理符号系统（PSS）的产物。符号是表示其他事物的模式;符号系统是一组符号，可以通过一组关系组成更复杂的结构。“物理”一词表达了符号系统可以而且必须以某种物理方式实现，以创造智能。物理基础可以是电子计算机的电路，思考生物有机体的神经基质，或者原则上可以实现图灵机之类的计算设备的任何其他东西。PSS假说暗示，结构化心理表征是人类智力的核心，在一段时间内没有争议，被大多数认知科学家接受为该领域的公理，几乎不需要理论分析或直接的经验支持。然而，在20世纪80年代中期，这一假说遭到了一些联结主义认知模型支持者的猛烈抨击，特别是那些“与许多其他模型平行”（参见Marcus，1997）的模型支持者。在这种模型中使用的表示通常被描述为“子符号”，因为基本单元对应于（相对）低级特征，在这些特征上，有意义的概念以分布式方式表示。只要基于“次符号”表征的模型实际上是非符号的，但足以作为人类智力的解释，那么对符号系统的需要就被消除了;因此，这一一般类别的模型构成了“消除性”联结主义（Pinker & Prince，1988）。消除联结主义对PSS假说提出了直接挑战，从而将后者从认知科学的公理转变为一个有争议的理论立场，无论基于分布式表征的模型是否为物理符号系统提供了真正的替代品，很明显，它们作为认知的可能算法账户具有吸引人的特性。离散符号以“全有或全无”的方式表示实体，从而违反了最小承诺原则（例如，用“狗”的存在或不存在来表示狗的存在或不存在，并不能提供直接依据来表达可能有狗的非决定性证据）。离散符号也不能表达所表示的实体的语义内容（例如，“狗”和“猫”的符号并不表示什么狗，
A foundational principle of modern cognitive science is the Physical Symbol System hypothesis, which states simply that human cognition is the product of a physical symbol system (PSS). A symbol is a pattern that denotes something else; a symbol system is a set of symbols that can be composed into more complex structures by a set of relations. The term " physical " conveys that a symbol system can and must be realized in some physical way in order to create intelligence. The physical basis may be the circuits of an electronic computer, the neural substrate of a thinking biological organism, or in principle anything else that could implement a Turing machine-like computing device. The PSS hypothesis, which implies that structured mental representations are central to human intelligence, was for some time uncontroversial, accepted by most cognitive scientists as an axiom of the field scarcely in need of either theoretical analysis or direct empirical support. In the mid-1980s, however, the hypothesis came under sharp attack from some proponents of connectionist models of cognition, particularly the advocates of models in the style of " parallel many others; see Marcus, 1997, for a review). The representations used in such models are often described as " subsymbolic " because the elementary units correspond to (relatively) low-level features, over which meaningful concepts are represented in a distributed fashion. In so far as models based on "subsymbolic" representations are actually non-symbolic, yet adequate as accounts of human intelligence, the need for symbol systems would be eliminated; hence models of this general class constitute "eliminative" connectionism (Pinker & Prince, 1988). Eliminative connectionism offers a direct challenge to the PSS hypothesis, thereby transforming the latter from an axiom of cognitive science into a controversial theoretical position, which has been vigorously Regardless of whether models based on distributed representations provide genuine alternatives to physical symbol systems, it is apparent that they have attractive properties as possible algorithmic accounts of cognition. Discrete symbols represent entities in an "all-or-none" fashion, thereby violating the principle of least commitment (e.g., using the presence or absence of the symbol "dog" to represent the presence or absence of a dog affords no direct basis for expressing inconclusive evidence that there may be a dog). Discrete symbols also fail to express the semantic content of the represented entities (e.g., the symbols "dog" and "cat" do not signify what dogs and