Page 35 - Deep Learning

18 Introduction

view of the world that is assumed in inductive learning theories. The details of
memory, induction, transfer and planning remain to be worked out, psycholo-
gists say, but the type of account sought is clear. Hundreds of experiments are
conducted every year in psychology departments around the world to fill in
those details. Empirical inductivism thus acts as an implicit meta-theory for
47
the study of cognitive change, a set of broad principles that do not by them-
selves explain any particular psychological phenomenon but specify what kind
of explanation psychologists should seek.
If reality – material as well as social – does not operate like a clockwork but
instead is complex, turbulent and unpredictable, empirical inductivism cannot be
the whole story of learning. Adapting to a turbulent world is not like figuring out
how a clock works. It is more like playing Meta-chess. This difficult game looks
exactly like an ordinary chess game. It is played with the same board and the same
pieces. unlike the case in ordinary chess, when a player moves a piece in Meta-
chess, he changes not only the location of that piece, but also the rules that control
how the pieces move. When a player moves a rook forward the rules change so
that pawns can only go backward, queens can only move three steps in any one
direction and knights travel only in straight lines. Each type of move changes the
rules in a different way. to select a move, a player has to think through not only
the changing relations between the pieces, but also how the associated changes in
the rules will affect the relative strengths of his own and his opponent’s positions.
to make matters worse, the set of rule changes associated with a move depends
on the number of moves made so far in the game. At intervals determined by the
decimal digits of the number pi, the rule changes change. The point is not that
this hypothetical game would be difficult to play, but that one could not learn to
play it via the processes envisioned in empirical inductivism. change is not con-
fined to the arrangement of pieces on the board, but reaches all the way down.
There are no stable features that could form the basis for an inductive theory of
the game. Strangely, Meta-chess is a well-defined game and yet turbulent through
and through.
If reality is more like Meta-chess than like a clock, if living is like playing
a game in which the rules change in ways that are themselves forever changing,
then this fact must be reflected in the cognitive mechanisms that evolved as
humans marched down the evolutionary pathway to cognitive flexibility. one
might object that discoveries about complex systems in the late 20th century
could hardly have affected pre-human hominids or archaic hunter-gatherers
150,000 years ago. Surely the self-organizing nature of climate change or the
possibility of a slowly drifting value for the gravitational constant would not
have troubled them.