opening it up with Common Lisp
Book review: Darwinia
Summer reading: Spin
the Omnivoire's Delimma
the Golem's Eye
A blast from the past (August, 1999!) review of Annette Karmiloff-Smith's wonderful Beyond Modularity:
Karmiloff-Smith proposes to view cognition and development as a series of Representational Redescriptions that occur across cognitive domains/modules and across developmental phases. Her view of mind is flexible and varied. She believes in a process of modularization (contra Fodor who claims that the modules are innately specified and not subject to change) and in a process of multiple domain specific phases (contra Piaget who claimed that development proceeds in domain general stages where the entire system changes at once from one stage to the next).
Representational Redescription (RR) is a model of how implicit procedural knowledge is encoded and re-encoded into more and more explicit forms until it finally becomes declarative. For example, when one learns to play a song at the piano, one first must play the piece as a whole. With time and practice, one becomes able to play parts of the piece without having the start each time at the beginning. Finally, one may be able to improvise with the piece and work with its parts in their own right. This movement from implicit "programmed" procedural knowledge to explicit declarative knowledge is re-enacted across domains and modules in multiple time scales and levels of detail.
The RR framework explains the ubiquitous U-shaped mastery curve: performance at a skill rises to a high level, then drops and then rises again. The RR framework explains this by saying that the initial mastery is due to an implicit understanding of the problem domain. Development continues after mastery is achieved, however, as redescription attempts to bring understanding to the implicit behaviors. The initial redescriptions often fail to properly distill the correct essence of the procedures and therefore produce less adequate performance. As redescription continues, a correct declarative model of the task is achieved and mastery returns.
Beyond Modularity is divided into an introduction; five chapters that discuss RR and development as it relates to Language, Physics, Mathematics, Psychology and Notation / Drawing; and two final chapters that discuss the more theoretical aspects of Karmiloff-Smith's work. Each chapter provides excellent reviews of relatively recent research into child development viewed through the lens of her RR framework. As an example, chapter 3: The Child as Physicist discuses an experiment in which 4-9-year olds were asked to balance blocks on a narrow support. Some blocks were normal, others had a weight glued to one end and still others had a weight hidden inside them. The 4- and 8-year-olds both perform well at the task regardless of the kind of block. 6-year-olds, however, continually attempt to balance the block at its midpoint, regardless of the kind of block that they are balancing. A more careful analysis of the children's behavior shows that 4-year-olds balance by proprioception alone whereas the 8-year-olds correctly classify each block-type and have an explicit understanding of how the weight affects the balance. 6-year-olds, however, have a model of balance that appears to be based entirely on length and their failure to balance the other blocks is viewed as anomalous data that can be rejected. Interestingly, the same 6-year-olds can easily use the very same blocks to build a house. It appears that there failure to balance only occurs when they are calling upon their explicit knowledge of balancing.
Beyond Modularity presents a flexible, engaging and non-dogmatic view of the development of mind. The RR framework appears to fit well with many of the observed behaviors of children and adults as they master new domains. That being said, the RR model remains silent as to how and why redescription actually occurs. What motivates redescription? Why are humans theoreticians and not just inductivists? How does the magic that turns implicit procedures into explicit theories function? Unless these questions can be answered, RR is an interesting implicit story of cognition that needs to undergo its own redescription into explicit form before it can be useful for actually building intelligent systems or truely understanding natural ones.
Copyright -- Gary Warren King, 2004 - 2006