Continuity and Discontinuity in Learning

Continuity and discontinuity in learning refer to the problem of whether there are qualitative changes in the fundamental mechanisms that govern learning. This question has been examined in the literature in two separate but related ways. The first approach examines whether qualitative breaks occur with increasing expertise or additional learning, and it was conducted mostly prior to the 1980s using animal learning paradigms and human learning of simple verbal materials. The second approach evaluates whether there are developmental changes in learning that are qualitative in nature.

The historical debate was inspired in part by the observation of seemingly contradictory phenomena. On one hand, much research in learning, including the seminal results from Hermann Ebbinghaus, revealed canonical smooth learning curves, in which performance could be plotted as a negatively accelerating approach to an asymptote. On the other hand, research by Robert Yerkes and Wolfgang Köhler using primates appeared to reveal ‘insight’ in basic problem-solving tasks. Complementing these reports [Page 186]were results from learning in rodents in which learning appeared to hover quite stably around chance levels of accuracy prior to rising rapidly to an asymptote. Even in tasks in which learning appears continuous, theorists such as William Estes and Randy Gallistel have shown that the learning functions for individual subjects may exhibit a more abrupt rise in performance than is comfortably accommodated within a view of learning that is purely continuous.

This dilemma motivated empirical and theoretical demonstrations from notable researchers like Kenneth Spence and Isadore Krechevsky (later David Krech) that continuous and gradual learning processes could underlie the apparently discontinuous learning curves that appeared to imply underlying learning processes that were more all-or-none than continuous. Such considerations probably played a large role in the sea-change in American psychology from stimulus-response theories to the mentalistic types of theorizing characterized by the work of Edward Tolman—theories that included previously taboo terms such as hypothesis generation and testing and insight, and that eventually became known as the cognitive revolution from the 1950s to the 1970s.

During that latter period, questions of continuity inspired one of the first widespread applications of basic mathematical models of psychological processes to questions about learning. The burgeoning set of researchers who pioneered the application of Markov processes to cognition, including William Estes, Patrick Suppes, James Greeno, and Gordon Bower, found that smooth learning curves could be produced by all-or-none learning functions, and that those all-or-none models accounted well for verbal learning data in humans. Two related empirical outcomes are central to the claim of supremacy for all-or-none models.

First, Estes showed that people tend not to recall items correctly on a second test when they have not correctly recalled them on a previous test. If discontinuous learning actually arose from variable levels of underlying learning, rather than the pure absence of learning, then an above-chance rate of recovery for previously unrecalled items would be expected. Second, there is no increase in memory performance with additional learning when only those trials prior to the final error for a particular item are examined. This, too, is consistent with the idea that those items reside in an as-yet unlearned state, rather than a state of gradually increasing learning that is nonetheless insufficient to support accurate performance.

...