348 Knowledge representations: networks

Aside re assessing good theory:  Properties of good theories

Explanatory power – can explain a particular behavior:

            Behavior A occurred because of Event/Situation B.

Predictive power – can predict results of particular experiments:

            Manipulation of IV = specific change in DV.

Falsifiability – theory can be tested to see whether it’s actually true.

Generation of experiments – theory stimulates research to test it,

            to improve it, to study new questions it raises.

Consider Freudian theory . . .

            Great explanatory power

            Weak predictive power

            Poor Falsifiability

            Poor generation of experiments

We’ll use the foregoing properties to evaluate network theories. 

They are also routinely used in science to evaluate all theories.

First question:  how is knowledge categorized?

            -- Classical/defining features approach

            -- Family resemblances approach

                        Prototypes

                        Exemplars

            -- Coherence/relations/roles approach

Second question:  How are relations among categories represented?

            --Semantic networks:  Collins & Quillian (1969), Collins & Loftus (1975)

            -- Connectionist models:  McClelland & Rumelhart (1986)

Semantic networks

            Nodes & links

            E.g., higher node (Animal) [skin – moves around – eats – breaths - . . .]

                        linked to lower node (Bird) [wings – can fly – feathers]

                                    linked to lowest node (Robin) [sings – red breast - . . .]

                        linked to lower node (fish) [fins – scales – gills – swims]

                                    linked to plecostomus [bottom-feeder – mustache!  . . .]

            Why not have all aspects of an instance (e.g., robin) stored with it?

            Cognitive economy

                        Why have “higher” and “lower” nodes?

                         – more efficient & hence faster to represent aspects of specific

                            instances as subsets of more general aspects

            Problem:  what about oddities, e.g., ostrich?

            Answer: node for the oddity would include relevant disclaimer.

                        For ostrich, the detailed disclaimer would be “can’t fly.”

            RT experiments confirm the foregoing:

                        We’re faster to say that a robin has a red breast

                        than to say that a robin has skin.

            Spreading activation

            Activation of a specific instance (e.g. robin) also activates the

            higher nodes – bird & animal

            This activation then spreads to other instances of bird (e.g., canary)

            Although we aren’t usually consciously thinking of the other

            instances, they are “primed” so it’s now easier & quicker for us to

            think of them if called upon to do so

            RT experiments in lexical decisions tasks support the priming prediction

                        [Task:  say yes/no to a list of items – words versus non-words]

            Problem:  typicality effects

                        RT for “A pig is an animal” faster than “A pig is a mammal”

                        Network says pig is a mammal is an animal,

                                    Thus pig = mammal should have been faster.

            Personal experience (Collins & Loftus)

            Abandon hierarchy and go with nodes & links based on personal experience

            This modification succeeded in eliminating the typicality effects & other problems

            In fact, it’s so effective that it explains essentially all categorization experiments

            Evaluation in terms of “Good theory” properties

                        Explanatory power –

                        Predictive power –

                        Falsifiability – [c.f. Johnson & Laird:  “Too powerful!]

                        Generation of experiments –

Connectionist models

            Input units – hidden units – output units

            An individual concept is represented in several units simultaneously

            Activations of units occur across several units simultaneously, “in parallel”

            Etc.

            Evaluation in terms of “Good theory” properties

                        Explanatory power –

                        Predictive power –

                        Falsifiability – [c.f. Johnson & Laird:  “Too powerful!]

                        Generation of experiments –