Trends in Cognitive Sciences
OpinionChunking mechanisms in human learning
Section snippets
Defining and observing chunks
The literature on chunking encompasses many different areas of research, and the concept of a chunk has consequently diversified in its meaning. The literature itself can be divided into two broad areas, based on how and when chunking is assumed to occur: the first assumes a deliberate, conscious control of the chunking process (goal-oriented chunking), and the second a more automatic and continuous process of chunking during perception (perceptual chunking). In spite of the surface variety
EPAM
Shortly after Miller's 1956 paper, Feigenbaum and Simon began to develop a pure and direct implementation of chunking mechanisms, known as EPAM (Elementary Perceiver and Memorizer) 13, 14, 15. Learning is simulated by the growth of a discrimination network, where internal nodes test for the presence of perceptual features, and leaf nodes store ‘images’, the internal representation of external objects. The learning mechanisms, detailed below, support the addition of information to leaf nodes and
CHREST
CHREST (Chunk Hierarchy and REtrieval STructures) 17, 27, 28, 29, 30 features a number of additions to EPAM's basic learning mechanisms, providing a greater degree of self-organization and adaptation to complex data. This section summarizes some of the new mechanisms within CHREST, before describing some applications.
Other computational approaches to chunking
Various computational approaches adopt the idea of chunking, either as a central or incidental feature. Chunking arises naturally within symbolic models of cognition, where elements of information are combined into single units. In this article, we have emphasized the use of a discrimination network to index long-term memory using the EPAM/CHREST family of models, although other such families exist 53, 54. An alternative approach relies on a production-rule representation for long-term memory.
Conclusion
This article has described and summarized chunking mechanisms in human learning, focusing on the EPAM/CHREST family of computational models and their applications. We have identified two broad classes of chunking: goal-oriented and perceptual chunking. From the diversity of available empirical evidence, the general notion of chunking appears to be a robust and important one in contemporary cognitive science. The lesson to be taken away from the EPAM/CHREST examples we have described is that
Questions for future research
- •
What is the neurophysiological plausibility of computational models based on chunking mechanisms?
- •
How can chunking mechanisms best be linked directly to sensors and effectors? And what impact will this have on the primitive elements on which chunks are based?
- •
How can perceptual and goal-oriented chunking mechanisms best be integrated?
- •
Can a theory based purely on chunking mechanisms lay any claims towards being a universal theory of cognition?
- •
Is language acquisition just like the acquisition of
Acknowledgements
The authors would like to thank Daniel Freudenthal, the anonymous referees, and the editor of Trends in Cognitive Sciences for their helpful comments on an earlier version of this article. This research was funded by the UK Economic and Social Research Council and the Leverhulme Trust.
References (60)
The magical number seven, plus or minus two: some limits on our capacity for processing information
Psychol. Rev.
(1956)- Cowan, N. (2001) The magical number 4 in short-term memory: a reconsideration of mental storage capacity. Behav. Brain...
- et al.
Skilled memory
- et al.
Expert chess memory: revisiting the chunking hypothesis
Memory
(1998) Structure and process in alphabetic retrieval
J. Exp. Psychol. Learn. Mem. Cognit.
(1983)- et al.
EPAM-like models of recognition and learning
Cognit. Sci.
(1984) The impact of chess research on cognitive science
Psychol. Res.
(1992)Simulation of expert memory with EPAM IV
Psychol. Rev.
(1995)A computer model of chess memory
Learning-based constraints on schemata