Page 251 - Deep Learning
P. 251

234                         Adaptation

            to coach someone else to perform better than oneself. It implies that when
            we adapt to new circumstances, we lead with our ability to evaluate outcomes
            rather than our ability to infer new conclusions.
               The second contribution of the constraint-based theory is the principle
            that practical knowledge starts out general and becomes more specific in the
            course of learning. There is a long-standing tradition, with roots in the begin-
            nings of Western philosophy, of viewing learning as moving in the opposite
                                                38
            direction, from particulars to abstractions.  Particulars are given in percep-
            tion while abstractions are human constructions, or so the ancient story goes.
            The hypothetical process that supposedly transforms particulars into abstrac-
            tions is called induction and it is often claimed to operate by extracting com-
            monalities across multiple particulars. If the first three swans you ever see are
            white, the idea swans are white is likely to come to mind. However, the notion
            of induction is riddled with problems. How are experiences grouped for the
            purpose of induction? That is, how does the brain know which experiences
            are instances of some abstraction X, before that abstraction has been learned?
            How many instances are needed? Which features are to be extracted? How are
            abstractions with no instances in human experience such as the infinite, the
            future and perfect justice acquired?
               The notion of induction has even less plausibility as an explanation for
            the acquisition of practical knowledge. We do not learn to tie our shoelaces by
            observing multiple instances of shoelace tying and extracting what they have
            in common. Skills are learned via practice, not induction, and what happens
            during practice is that initial, vague and underspecified action tendencies and
            dispositions become better adapted to a task environment by incorporating
            information from that environment. The end point of practice is not abstrac-
            tion but competence, a collection of tactics for operating effectively in a par-
            ticular environment. According to the specialization principle, the acquisition
            of practical knowledge thus moves in the opposite direction to what is often
                                                       39
            assumed in education, philosophy and psychology.  The concept of speciali-
            zation resonates better with basic themes in the biological sciences: Both evo-
            lution and ontogenesis proceed from general, undifferentiated structures to
            more differentiated and specific ones.
               Third, the theory focuses attention on the information in negative rather
            than positive outcomes. It is natural to expect a target strategy to be learned
            primarily from correct and positive information. Indeed, it is an axiom in
            Western culture in general and the culture of U.S. education in particular that
            positive feedback is more helpful than negative feedback. This is undoubtedly
            so when feedback is interpreted as praise or punishment. But it is not obvious
   246   247   248   249   250   251   252   253   254   255   256