Why emergentist accounts of cognition are more theoretically constraining than structured probability accounts: comment on Griffiths et al. and McClelland et al.
Why emergentist accounts of cognition are more theoretically constraining than structured probability accounts: comment on Griffiths et al. and McClelland et al.
复制标题
为什么认知的涌现论解释比结构化概率解释在理论上更具约束力：对 Griffiths 等人的评论。
DOI：
10.1016/j.tics.2010.05.013
复制
发表时间：
2010
影响因子：
19.9
通讯作者：
Altmann GT
中科院分区：
文献类型：
--
作者：
Altmann GT
认知的结构化概率方法旨在通过识别其所面临的归纳问题的理想解决方案来逆向工程人类思维[1]。动态系统方法[2]将认知行为解释为学习环境中统计学假设的紧急结果。这两种方法都提供了关于环境中的统计规律如何映射到人类认知的见解。然而，就其本身而言，理想解决方案的方法类似于将进化建模为一个“心中”有目标的过程，因此必须找到一个达到该目标的解决方案，而不管环境对目标是否可以通过支持进化的渐进（自然）选择过程实现的限制。动力系统方法考虑了“外源性”环境约束和对生物体成熟状态的动态变化的“内源性”约束，以解释最终状态如何受到这些不同“压力”的约束[3]。在这样做的过程中，它也解释了两个相关的现象，迄今在很大程度上被结构化概率方法所忽视：跨时间采样环境的后果[4]和对象表征与概率归纳之间的关系。一个已经得到很好研究的涌现系统是简单的递归网络[5];它以固定的时间分辨率对输入进行采样，但原则上能够允许出现跨越多个时间分辨率的表示（即捕捉不同类型的偶然性，每个偶然性都跨越不同的时间框架，在网络所暴露的连续输入之间）[4]。实际上，出现了等级代表（反映了不同时间框架内发生的系统变化）。因此，通过一种抽样机制的运作，成年人能够在多个时间框架内对分层表征和多种偶然事件进行编码，这种机制最初可能只局限于一个时间框架。
Structured probability approaches to cognition aim to reverse-engineer the human mind by identifying ideal solutions to the inductive problems that it faces [1]. Dynamical systems approaches [2] explain cognitive behaviour as the emergent consequence of statistical regularities in the learning environment. Both approaches offer insights into how statistical regularity in the environment can map onto human cognition. However, taken on its own, the ideal solution approach is akin to modelling evolution as a process that has a goal ‘in mind’, whereby a solution through which to arrive at that goal must be found regardless of the environmental constraints on whether the goal is achievable through the gradual (natural) selection processes that underpin evolution. The dynamical systems approach takes account of both ‘exogenous’ environmental constraints and (dynamically changing)‘endogenous’ constraints on an organism’s maturational state to explain how the final state is constrained by these different ‘pressures’[3]. In doing so, it also explains two related phenomena that have hitherto been largely overlooked by structured probability approaches: the consequences of sampling the environment across time [4] and the relationship between object representation and probability induction.One emergentist system that has been well studied is the simple recurrent network [5]; it samples the input with a fixed temporal resolution but is in principle able to allow the emergence of representations that span multiple temporal resolutions (ie that capture different kinds of contingency, each across different time frames, between the successive inputs to which the network is exposed)[4]. In effect, hierarchical representation emerges (reflecting systematic variation occurring across different time frames). The adult ability to encode hierarchical representations and multiple contingences across multiple time frames could therefore emerge through the operation of a sampling mechanism that might initially be bound to just a