Page 62 - Deep Learning
P. 62

The Nature of the Enterprise             45

            that such-and-such a learning mechanism was operating on such-and-such an
            occasion is an ad hoc reconstruction instead of an explanatory account.


                               Patterns as Explanatory Targets
            Sometimes the target for a cognitive explanation is a single learning event, a
            change in knowledge that a person underwent on a particular occasion. For
            example, we might want to explain why a student failed to learn a piece of sub-
            ject matter in a given learning scenario. This type of detailed analysis of single
            events has precedents in other sciences: Evolutionary biologists might attempt
            to explain the evolution of a particular species, while economic analysts might
            try to explain the rise and fall of a particular company. 49
               However, the more significant explanatory targets are patterns of change.
            A pattern is not a theory or an explanation, but a description of the general
            character of the change a system undergoes over time. To take an example
            from paleontology, the temporal pattern in the fossil record is one of relatively
            short periods of speciation followed by much longer periods of stasis during
            which a species does not change, which in turn are followed by relatively short
            periods of decline, ending in extinction; this pattern is known as punctuated
                      50
            equilibrium.  In economics, another well-known pattern is the existence of
            business cycles, the regular succession of booms and busts in the market. In the
            Earth sciences, the pattern of advance and retreat of the world’s glaciers is a
            major explanatory target.  In human cognition, the best-documented pattern
                                 51
            of change is the learning curve: If the time to complete a practice problem is
            plotted as a function of how many practice problems the person has solved
            already, the result is a smoothly decelerating curve.
               Patterns are important in part because they are more difficult to explain
            than individual learning events. Any single learning event can be explained if
            we are free to include any representation and any repertoire of basic processes
            among our premises. But an explanation for a pattern must show not only how
            the change can come about but also why it recurs. If geologists had found only
            a single meandering river on the entire planet, they could have attributed it to
            unique local conditions, but because rivers meander all over the Earth, the expla-
            nation for meandering draws upon recurring factors such as silt deposit and bank
                  52
            erosion.  Patterns are powerful clues to the mechanisms that produce them.
               Patterns of change often unfold over a period of time that is significantly
            longer than the duration of a single unit change. A pattern is a cumulative
            outcome of multiple applications of the basic change mechanisms. A focus on
            patterns forces the theorist to consider how the effects of the basic mechanisms
   57   58   59   60   61   62   63   64   65   66   67