Page 33 - Deep Learning
P. 33
16 Introduction
change that was too rapid for this type of adaptation. Learners, on the other
hand, can deal even with rapid change.
our superior ability to learn is arguably our species-specific character-
istic, so a theory of the human mind requires a learning theory as one of its
components. As long as psychologists conceptualized the environment within
the clockwork mind-set, it was clear in principle how learning works: Because
change is an illusion and reality consists of stable regularities, a person can
learn by accumulating experiences, analyze them to identify the regularities,
project those regularities onto the future and act accordingly.
This requires certain cognitive capabilities. to accumulate experience
requires memory. Each event or experience must leave a trace of some sort –
richard W. Semon suggested that we call it an engram but it is now more often
called an episodic memory – which can be reactivated later to inform action at
41
some future time. Although early efforts by Karl Lashley and others to locate
individual engrams in the brain were unsuccessful, the existence of such traces
is not in doubt. Every person can recall any number of events and situations
42
that have happened in his own past, and textbooks in cognitive psychology
require multiple chapters to describe the many experiments that investigate
our capacity to form, retain and retrieve episodic memories. 43
Memories of individual events are not very useful in themselves, but,
according to the received view, they form the raw material for further learn-
ing. By extracting the commonalities across a set of related episodic memories,
we can identify the underlying regularity, a process variously referred to as
abstraction, generalization or induction. The output of this supposed process
is a knowledge structure that summarizes a regularity that was present in a
person’s past experience. There is little doubt that people are capable of iden-
44
tifying patterns in sequences of experiences. People obviously form general
concepts like food, kin and tool, and the march of the seasons, the phases of
the moon and the cycles of the tide must have been identified by early humans
long before the systematic investigations we now call science began.
In the clockwork view, general knowledge is useful because the regularities
in past experience are also the regularities that will control future experiences. We
can form more or less precise expectations about what future situations will be
like on the basis of what the past was like. Indeed, the past is our only source of
information about the future. The process of applying a regularity from the past to
future situations is traditionally called transfer by psychologists and induction by
logicians, but the philosopher nelson Goodman suggested the more descriptive
term projection. under any name, this process is useful to the extent that regulari-
45
ties in past experience continue to operate in the future.