Page 382 - Deep Learning

Elements of a Unified Theory 365

of human cognition? Behavior differs from situation to situation and from per­
son to person, and it is obviously a function of a person’s prior learning history,
which in turn is infinitely variable. if so, what kind of statement can serve as
a general principle in cognitive psychology? Psychologists with a reductionist
mind­set assume that analysis can be pursued all the way down to brain mod­
ules and neurons, subsuming mental function under whatever generalizations
9
hold for neural tissue. They overlook the fact that the reduction of mind to
brain must cross the mind­brain chasm along the way, and nobody knows
what kind of bridge might serve. The majority of cognitive psychologists pur­
sue the analytical strategy but stop short of neural reductionism: High­level
cognitive functions like problem solving, memory and decision making are
explained through analysis, not into neurons, but into combinations of more
basic cognitive processes: Associate two previously unconnected concepts,
spread activation from one memory node to another, bind a variable to a par­
ticular entity and so on. newell has suggested that an analytical level of cog­
10
nition can be specified in terms of its time band, the range of the durations of
the constituent processes. newell distinguished between four time bands, but
11
most cognitive psychologists make do with two: The higher­level functions
typically studied in psychological experiments execute in minutes or hours,
while the basic processes postulated in the explanations of those functions are
claimed to execute in seconds or fractions of seconds.
After the cognitive revolution in the late 1950s, it was not yet clear how
analysis into basic cognitive processes might serve unification. Researchers
published their own flow diagram of whatever cognitive process caught their
interest. After two decades of this, it was clear that this research practice was
12
not converging on a general theory.
in two remarkable papers published in the early 1970s, newell proposed
a type of theory that slices cognition into two distinct parts: the infinitely var­
iable and constantly changing knowledge base and the stable machinery that
utilizes that knowledge to perform tasks. Although newell called the second
13
part the “control structure” of the mind, a decade later J. R. Anderson pro­
posed the less engineering­sounding label “cognitive architecture” and the
latter caught on. The description of the cognitive architecture can be concep­
tualized as a blueprint for a particular type of computer, namely, the type of
which a human brain is an instance. in this analogy, knowledge plays the role
of software. The specification of a cognitive architecture is a classical analytical
enterprise: The system as a whole, human cognition, is understood by breaking
it down into its parts, and each part into its parts, and so on. The underlying
assumption is that the basic machinery of cognition is at least approximately