13.+Concepts+and+Categories

April 7, 2015
 * 13. Concepts and Categories**

How concepts are represented Learning concepts
 * Outline**
 * Prototypes, exemplars, and theories
 * Semantic and connectionist networks

Categorization
 * Prototypes, exemplars, and theories**
 * We categorize new things by comparing them to our mental representation of the category
 * How is that representation stored?
 * Prototype theory - store just the prototype
 * Problems with prototype theory by itself:
 * There's no average (ex: instruments)
 * Prototypes don't capture our knowledge of category variability (ex: there are many varieties of mustard but only one variety of ketchup)
 * Exemplars[[image:more dogs.png width="168" height="130" align="right"]]


 * Evidence suggests we use both prototypes and exemplars


 * Semantic networks**
 * Spreading activation: activation spreads from activated nodes to related nodes
 * Spreading activation leads to //simultaneous multiple constraint satisfaction//
 * Multiple cues/constraints work simultaneously to elicit the target
 * Examples: men's names that sound like verbs
 * Remote Associates Test (RAT) - finding associations to a fourth item making the item activate
 * Simultaneous multiple constraint satisfaction
 * Example: Cracker Union Rabbit (Jack); Bald Screech Emblem (Eagle), etc.
 * More examples to try:
 * Inch Deal Peg
 * Jump Kill Bliss
 * Magic Plush Floor
 * Note Dive Chair
 * answers at bottom of page
 * Also used as a measure of creativity

Three lines of evidence:
 * Evidence for Semantic Networks**
 * Semantic priming
 * Semantic distance effect
 * Typicality effects


 * Semantic priming**
 * Meyer & Schvaneveldt (1971): Subjects see two strings of letters, do a lexical decision task. Told to respond "yes" if both are words, "no" if they're not.
 * Example:
 * Vault-house
 * Page-tall
 * Two conditions of interest:
 * Related words: nurse-doctor
 * Unrelated words: bread-doctor
 * Results: Faster reaction for related word pairs. Items prime semantically related items.
 * [[image:semantic priming.png width="443" height="289"]]
 * Semantic distance effect**: Categorization and trait verification take longer with each additional step in the semantic network
 * Example: Take longer to verify that a canary has skin than that a canary sings, because more steps are required to get there.
 * A canary sings: Canary --> Sings
 * A canary has skin: Canary --> bird --> animal --> skin


 * Typicality Effects**: faster categorization for typical members
 * Smith, Rips, & Shoben (1974)
 * Sentence verification task
 * "A penguin is a bird" was slower
 * "A robin is a bird" was faster
 * Why:
 * Not because of the number of steps...
 * The robin-bird link is stronger than the penguin-bird link

Semantic networks are similar to neural networks Differences
 * Connectionist Networks**
 * Links
 * Nodes
 * Spreading activation
 * All or none firing
 * In semantic networks, each node has meaning
 * In neural networks, nodes have no meaning in isolation
 * Which brings us to connectionist models

Connectionist models: The pattern of activation (i.e., combination of nodes) represents the concept
 * Like a semantic network:
 * models semantic knowledge (e.g., "bird")
 * Like a neural network:
 * Individual nodes do not have meaning
 * AKA parallel distributed processing (PDP)
 * Spreading activation happens in parallel
 * Concept representation ins distributed across nodes
 * In a semantic network, ice cream flavors would look like this:
 * [[image:semantic ice cream.png width="218" height="166"]]
 * In a connectionist network, ice cream flavors would look like this
 * [[image:chocolate.png width="105" height="82"]][[image:vanilla.png width="100" height="83"]][[image:strawberry.png width="104" height="83"]]


 * Category Learning**
 * Inductive learning: the ability to generalize concepts through exposure to multiple exemplars
 * Effect of spacing on inductive learning:
 * Spacing makes it difficult to identify a category's unifying characteristics
 * Massing makes doing so much easier
 * Experiment: Different artists' work presented 6 times each during study.
 * A given artist's paintings are either massed or spaced.
 * Result: Spaced condition performed better
 * [[image:inductive learning.png width="270" height="227"]]
 * Participants asked whether massed or spaced helped them learn more
 * [[image:massed or spaced.png width="417" height="196"]]
 * Summary:
 * Learning a concept requires noticing...
 * Similarities within a category
 * Differences between categories
 * Spacing is a desirable difficulty
 * Spacing enhances long-term learning
 * Spacing impairs performance during study, so an effective strategy seems counterproductive

Answers to remote association test: Square Joy Carpet High