Page 387 - Deep Learning
P. 387

370                         Conclusion


                           structured, unbounded Representations

                                                                           17
            Mental representations exhibit what psycholinguists call constituent structure.
            A representation consists of components, parts, which themselves are mental
            representations. Every mental representation at layer N is both a combination
            of constituents from layer N­1 and itself a constituent of some representation
            of greater complexity or scope at layer N+1. The number of layers cannot be
            determined with any certainty or precision and is likely to vary from one type
            of representation to another. in perception, many psychological accounts oper­
            ate with only two layers, visual features and objects, but vision scientists have
            identified perhaps a dozen distinct layers: features, contours, objects and scenes,
            with further divisions within each.  in the study of discourse comprehension,
                                        18
            researchers distinguish between concepts, propositions, text bases and situation
            models.  in the study of cognitive skills, researchers distinguish between actions,
                  19
                                                                          21
            individual rules and strategies,  and in the study of declarative knowledge,  it
                                     20
            is necessary to distinguish between concepts, beliefs and belief systems. it mat­
            ters how the components of a representation are combined. The structure of
            relations among the components is not imposed on the representation, but is an
            integral part of the representation. A representation at layer N can be changed by
            selecting different elements at layer N­1 or by combining them differently.
               A repertoire of elements at layer N in conjunction with the rules for how
            they combine to generate a possibility space at level N+1, and the astronomi­
            cal  sizes  of  such  combinatorial  spaces  guarantee  that  surprises  will  never
            cease. As an example from the material world, the number of different carbon
                                         22
            compounds runs into the millions.  With respect to the living world, biolo­
            gists report that there are more ways of making a beetle than one would have
            thought; taxonomists recognize some 350,000 different species, and that num­
            ber can only increase over time as evolution throws new species down and
                                23
            naturalists pick them up.  in the domain of human creativity, we can ask how
            many different ways there are to put paint on a canvas: Beetles and painters
            might be neck and neck with respect to the sizes of their respective possibil­
            ity spaces. The concept of combination goes a long way toward explaining the
            diversity and richness of mental representations.
               By itself, the concept of a combinatorial space commits us to the counterin­
            tuitive conclusion that cognition is bounded in the sense that there is a large but
            finite and predetermined space of possibilities that we cannot transcend. To reach
            deeper, we have to recall that the edge of a possibility space is movable, because
            changes in the underlying layer might generate a different set of elements or new
            ways of combining them. The projection from layer N to layer N+1 creates the
   382   383   384   385   386   387   388   389   390   391   392