Page 218 - Deep Learning
P. 218

The Growth of Competence                201

               The fact that there are multiple theories and models of skill acquisition
            should not fool us into conducting experiments with the purpose of selecting
            one of them as the better theory. Competitive comparisons of goodness of
            fit to experimental data are only meaningful when two hypotheses are mutu-
            ally incompatible in the strong sense that if one is true the other cannot be.
            But this condition does not hold for hypotheses about individual skill acquisi-
            tion mechanisms because nothing prevents human beings from learning in
            multiple ways. The theoretical task is not to pit different learning mechanisms
            against each other in a race to account for this or that phenomenon, but to
            assemble a repertoire of mechanisms such that we can explain as many aspects
            of skill acquisition and skill adaptation as possible.
               Human beings are efficient learners, and reliance on a protracted period of
            learning is a species-characteristic trait, so it is plausible that we have evolved to
            be maximally efficient learners. As our eyes are maximally efficient in the sense
            that the retina can react to a single photon, so our ability to learn might have
            evolved to make use of every bit of information.  to be maximally efficient
                                                     66
            is to make maximal use of all information that might be available to a learner
            during practice. This Principle of Maximally Efficient Learning is an example of
            what cognitive scientists call a rationality principle.  it is an idealization, but
                                                       67
            it is close enough to the truth to anchor the search for an empirically adequate
            theory. The implication is that to propose a theory of skill acquisition is to
            assemble a repertoire of nine different learning mechanisms to make use of the
            nine types of information.
               The  pool  of  hypotheses  proposed  and  formalized  in  simulation  mod-
            els  already  contains  one  or  more  well-specified  mechanisms  to  make  use
            of each of the nine information types. The surprising conclusion is that we
            already possess a first-approximation, multimechanism theory of skill acquisi-
            tion. More precisely, we possess a family of such theories in which the mem-
            bers vary slightly, depending on exactly which variant of each mechanism we
            select from the smorgasbord of proposals. Abstracting over those details for
            a moment, the nine Modes Theory claims that people acquire skills through
            nine different learning mechanisms: (a) Proceduralization, or translating ver-
            bal instructions into executable practical knowledge, perhaps in the form of a
            set of production rules. (b) Practical reasoning, or reasoning from prior declar-
            ative knowledge about the task environment to a conclusion about what to do.
            (c) Analogy, or mapping the current task onto an analogical task in memory
            and using the strategy for the analogue as a template for the strategy-to-be-
            learned. (d) Internalization of demonstrations or solved examples. (e) Caching
            outcomes of unselective search (trial and error). (f) Generalization of tentative
   213   214   215   216   217   218   219   220   221   222   223