Ohayon M
Ann Med Psychol (Paris). 1990 Oct;148(8):669-95.
It is clear that computers are but a poor brain models: the nervous system has many "processors" (neurons) in parallel, whereas von Neuman's machines work sequentially on a single processor. In complex systems, emergent properties cannot be inferred from the behaviour of single elements. Anthills display collective "meaningful" moves, while each ant seems to obey local interactions only. Likewise, large parallel networks of processing elements elicit emergent properties. Like brains, some of them are self-organizing systems. In large parallel processing networks, each unit performs an elementary computation: adding inputs from other units. Large nets display surprising spontaneous computational abilities: associative memories, classes, generalizations may be seen as emergent properties of the network. Symbols are dynamical entities, whose handing is driven by local interactions of activation/inhibition of related representations. In such models, representations (memories) are distributed in the whole network, as stable configurations. Indeed, the basic properties of representation in connectionist models seem closer to human mental objects than the classic Artificial Intelligence concepts. Connectionist models have been used in many fields, namely simulations of real neural networks, pattern recognition and artificial vision, speech recognition, language understanding and knowledge representation, problem solving... Connectionist models have been thus used in neurobiology as well as cognition. One basic structure seems indeed able to account for a range of cognitive functions, from perception to problem solving and high level cognitive tasks. Nevertheless studies about "pathological" networks are yet rare, still an open field... We explore some of these fields.
很明显，计算机不过是一种糟糕的大脑模型：神经系统有许多并行的“处理器”（神经元），而冯·诺依曼机器在单个处理器上顺序工作。在复杂系统中，涌现特性无法从单个元素的行为推断出来。蚁丘展现出集体的“有意义”行动，而每只蚂蚁似乎只遵循局部的相互作用。同样，大量并行的处理元件网络会引发涌现特性。其中一些像大脑一样，是自组织系统。在大规模并行处理网络中，每个单元执行基本的计算：将来自其他单元的输入相加。大型网络展现出惊人的自发计算能力：联想记忆、分类、归纳可以被视为网络的涌现特性。符号是动态实体，其处理由相关表征的激活/抑制的局部相互作用驱动。在这样的模型中，表征（记忆）作为稳定的构型分布在整个网络中。事实上，与经典人工智能概念相比，联结主义模型中表征的基本特性似乎更接近人类心理对象。联结主义模型已被应用于许多领域，即真实神经网络的模拟、模式识别与人工视觉、语音识别、语言理解与知识表征、问题解决……联结主义模型因此也被用于神经生物学以及认知研究。一种基本结构似乎确实能够解释一系列认知功能，从感知到问题解决以及高级认知任务。然而，关于“病理”网络的研究仍然很少，仍是一个开放的领域……我们探索其中一些领域。
