NOTE: DIFFERENCES BETWEEN A PHYSICAL SYMBOL SYSTEM AND A CONNECTIOMST (PDP) ARCHITECTURE

1. LOCAL VERSUS DISTRIBUTED REPRESENTATIONS
-IN A CONNECTIONIST MODEL THE RELEVANT IS NOT REPRESENTED LOCALLY, SUCH AS "NODES" THAT RESEMBLE "CONCEPTS" (IN A CONNECTIONIST MODEL THERE IS NO NODE FOR A CONCEPT SUCH AS RIPPlE, PRIEST, PARK ETC)

RATHER INFORMATION IS DISTRIBUTED OVER A NUMBER OF PROCESSING UNITS, NONE OF WHICH STAND FOR A CONCET
 

2. IN A CONNECTIONIST MODEL, INFORMATION IS CONTAINED IN THE WEIGHTS BETWEEN PROCESSING UNITS; INFORMATION DEPENDS ON A PAHERN OF ACTIVATION; THE SAME UNITS
MIGHT BE INVOLVED IN QUITE DIFFERENT INFORMATION

3. CONNECTIONIST MODELS LEARN VIA EXPERIENCE.. .AND DO SO BY CHANGING THE WEIGHTS BETWEEN UNITS


AN EARLY EXAMPLE OF A CONNECTIONIST MODEL.. BUILT TO RECOGNIZE PATTERN (NOTE THE DIFFERENCE WITH A SYMBOLIC MODEL, SUCH AS "PANDEMONIUM')

units have threshold that must be met before they "fire"
fire=either inhibitory or excitatory
magnitude of effect depends on weighting of the link

SO: in example above, output unit will fire if either features is present (feature designated here as either "1" or "0")

System LEARNS; if it responds that a pattern is present when in fact it isn't, then a reduction is made in the strengths of connections from all units that are currently active

-if it fails to recognize a presented pattern, connection strengths are increased

NOTE: some problems: the model will fire if either feature is present but also if both are present

By 1969, some argued (on mathematical analyses) limitations of the sort mentioned above were too severe and in principle unsolvable.. .they concluded that one should build models using physical symbol system architectures


LATER WORK HAS OVERCOME MANY OF THESE SO-CALLED UNSOLVABLE PROBLEMS BY ADDING LAYERS OF "HIDDEN UNITS" (IF UNITS NOT DIRECTLY RELATED TO INPUT OR TO PROVIDE OUTPUT)

a problem was how learning occurs
ONE SOLUTION: BACKWARD PROPOGATION

Psych 235 Home Page