[quoting Loser, 14 Apr 95] Structuralism has been applied to resolve scientific materialism-idealism conflicts. I have found no similar attempts to reconcile behaviorist-constructivist conflicts. Is any subscriber to ITForum aware of such an effort?
Some recent cognitive science work in machine learning supports this. The following abstract from a special session after the Australian Cognitive Science Conference suggests that we benefit from some pre-processing to abstract information before we do learning on it. I might also argue that the connectionist/sub-symbolic approach in cognitive science is also viewed by some as a way around the constructivist-behaviourist conflict, though not specifically for educational purposes.
Although I have to say, in the structuralist solution, the question then seems to reside in where you draw the line between mind constructing and mind reflecting; if the pre-processing happens "in the head" then this would argue for the idealist side, wouldn't it?
Chris Thornton
School of Cognitive and Computer Sciences
University of Sussex
Over the past 40 years or so a huge number of machine learning methods have been developed. These are of interest to the computer engineer because they offer a way of obtaining "machines" (computers) without specifying exactly how they should work (be programmed). They are also of relevance to the task of explaining natural cognition since they provide plausible models for natural learning processes. Unfortunately, it is only too clear that the performance of machine learning methods is greatly inferior to that of even quite primitive natural organisms. The cash-value that these methods have for both science and engineering is therefore somewhat limited.
Recently some researchers have begun to realize that the performance of learning methods can often be improved by recoding the data upon which the learning is performed. A learner can thus improve its own performance by decomposing the learning task it is confronted with into (1) a recoding task and (2) a simpler learning task. This strategy is simply an application of the old idea of "divide-and-conquer." However, when we come to look at the situation more closely we find that recodings which are good for learning are precisely those which provide explicit representations for abstract properties of the learner's environment. This is an interesting outcome. One implication is that advanced learning requires the formation of abstract representational structure. Another is that sophisticated learners (e.g., humans) will tend to use abstract representations of their environments.