Chapter 9 Knowledge. Some Questions to Consider Why is it difficult to decide if a particular object belongs to a particular category, such as “chair,”

Chapter 9

Knowledge

Some Questions to Consider

• Why is it difficult to decide if a particular object belongs to a particular category, such as “chair,” by looking up its definition?

• How are the properties of various objects “filed away” in the mind?

• How is information about different categories stored in the brain?

• Can young infants respond to the categories “cat” and “dog”?

Knowledge

• Concept: mental representation used for a variety of cognitive functions

• Categorization is the process by which things are placed into groups called categories

Why Categories Are Useful

• Help to understand individual cases not previously encountered

• “Pointers to knowledge”

– Categories provide a wealth of general information about an item

– Allow us to identify the special characteristics of a particular item

Caption: Knowing that something is in a category provides a great deal of information about it.

Definitional Approach to Categorization

• Determine category membership based on whether the object meets the definition of the category

• Does not work well

• Not all members of everyday categories have the same defining features

Caption: Different objects, all possible “chairs.”

Definitional Approach to Categorization

• Family resemblance

– Things in a category resemble one another in a number of ways

The Prototype Approach

• Prototype = “typical”

• An abstract representation of the “typical” member of a category

• Characteristic features that describe what members of that concept are like

• An average of category members encountered in the past

• Contains the most salient features

• True of most instances of that category

Caption: Results of Rosch’s (1975a) experiment, in which participants judged objects on a scale of 1 (good example of a category) to 7 (poor example): (a) ratings for birds; (b) ratings for furniture.

The Prototype Approach

• High-prototypicality: category member closely resembles category prototype

– “Typical” member

– For category “bird” = robin

The Prototype Approach

• Low-prototypicality: category member does not closely resemble category prototype

– For category “bird” = penguin

The Prototype Approach

• Strong positive relationship between prototypicality and family resemblance

• When items have a large amount of overlap with characteristics of other items in the category, the family resemblance of these items is high

• Low overlap = low family resemblance

The Prototype Approach

• Typicality effect: prototypical objects are processed preferentially

– Highly prototypical objects judged more rapidly

• Sentence verification technique

Caption: Results of E.E. Smith et al.’s (1974) sentence verification experiment. Reaction times were faster for objects rated higher in prototypicality

The Prototype Approach

• Typicality effect: prototypical objects are processed preferentially

– Prototypical objects are named more rapidly

The Prototype Approach

• Typicality effect: prototypical objects are processed preferentially

– Prototypical category members are more affected by a priming stimulus

– Rosch (1975b)

• Hearing “green” primes a highly prototypical “green”

Caption: Procedure for Rosch’s (1975b) priming experiment. Results for the conditions when the test colors were the same are shown on the right. (a) The person’s “green” prototype matches the good green, but (b) is a poor match for the light green.

The Exemplar Approach

• Concept is represented by multiple examples (rather than a single prototype)

• Examples are actual category members (not abstract averages)

• To categorize, compare the new item to stored examples

The Exemplar Approach

• Similar to prototype view– Representing a category is not defining it

• Different: representation is not abstract– Descriptions of specific examples

• The more similar a specific exemplar is to a known category member, the faster it will be categorized

The Exemplar Approach

• Explains typicality effect

• Easily takes into account atypical cases

• Easily deals with variable categories

Prototypes or Exemplars?

• May use both

• Exemplars may work best for small categories

• Prototypes may work best for larger categories

Caption: Levels of categories for (a) furniture and (b) vehicles. Rosch provided evidence for the idea that the basic level is “psychologically privileged.”

Caption: Left column: category levels; middle column: examples of each level for furniture; right column: average number of common features, listed from Rosch, Mervis et al.’s (1976) experiment.

Evidence that Basic-Level Is Special

• People almost exclusively use basic-level names in free-naming tasks

• Quicker to identify basic-level category member as a member of a category

• Children learn basic-level concepts sooner than other levels

• Basic-level is much more common in adult discourse than names for superordinate categories

• Different cultures tend to use the same basic-level categories, at least for living things

A Hierarchical Organization

• To fully understand how people categorize objects, one must consider

– Properties of objects

– Learning and experience of perceivers

Caption: Results of Tanaka and Taylor’s (1991) “expert” experiment. Experts (left pair of bars) used more specific categories to name birds, whereas nonexperts (right pair of bars) used more basic categories.

Semantic Networks

• Concepts are arranged in networks that represent the way concepts are organized in the mind

Semantic Networks

• Collins and Quillian (1969)

– Node = category/concept

– Concepts are linked

– Model for how concepts and properties are associated in the mind

Caption: Collins and Quillian’s (1969) semantic network. Specific concepts are indicated in blue. Properties of concepts are indicated at the nodes for each concept. Additional properties of a concept can be determined by moving up the network, along the lines connecting the concepts. For example, moving from “canary” up to “bird” indicates that canaries have feathers and wings and can fly.

Semantic Networks

• Cognitive economy: shared properties are only stored at higher-level nodes

• Exceptions are stored at lower nodes

• Inheritance– Lower-level items share properties of higher-level

items

Caption: The distance between concepts predicts how long it takes to retrieve information about concepts as measured by the sentence verification technique. Because it is necessary to travel on two links to get from canary to animal (left), but on only one to get from canary to bird (right) it should take longer to verify the statement “a canary is an animal.”

Semantic Networks

• Spreading activation– Activation is the arousal level of a node– When a node is activated, activity spreads

out along all connected links– Concepts that receive activation are

primed and more easily accessed from memory

Semantic Networks

• Lexical decision task

– Participants read stimuli and are asked to say as quickly as possible whether the item is a word or not

Semantic Networks

• Myer and Schvaneveldt (1971)

– “Yes” if both strings are words; “no” if not

– Some pairs were closely associated

– Reaction time was faster for those pairs

• Spreading activation

Semantic Networks

• Criticism of Collins and Quillian

– Cannot explain typicality effects

– Cognitive economy?

– Some sentence-verification results are problematic for the model

Semantic Networks

• Collins and Loftus (1975) modifications

– Shorter links to connect closely related concepts

– Longer linkers for less closely related concepts

– No hierarchical structure; based on person’s experience

Caption: Semantic network proposed by Collins and Loftus (1975). (Reprinted from A.M. Collins & E.F. Loftus, “A Spreading-Activation Theory of Semantic Processing.” From Psychological Review, 82, pp. 407-428, Fig. 1. Copyright © 1975 with permission from the American Psychological Association.

Assessment of Semantic Networks

• Is predictive and explanatory of some results, but not all

• Generated multiple experiments

• Lack of falsifiability

– No rules for determining link length or how long activation will spread

– Therefore, there is no experiment that would “prove it wrong”

– Circular reasoning

The Connectionist Approach

• How do neurons that can only either signal or not signal (on-off) represent knowledge?

• Model on digital circuitry – 1 bit can only be on or off– Group them together and get many more

patterns (1 byte: 00110011)

• Our neurons group together to form a network to represent concepts

• A network of nodes and links but operates very differently from semantic networks

The Connectionist Approach

• “Neuron-like units”

– Input units: activated by stimulation from environment

– Hidden units: receive input from input units

– Output units: receive input from hidden units

Caption: A connectionist network showing input units, hidden units, and output units. Incoming stimuli activate the input units, and signals travel through network, activating the hidden and output units. Note that this is an extremely simplified version of a connectionist network. Networks used in research on connectionism contain many more units and more complex connections between units.

The Connectionist Approach

• Parallel distributed processing

• Knowledge represented in the distributed activity of many units

• Weights determine at each connection how strongly an incoming signal will activate the next unit

The Connectionist Approach

• How learning occurs

– Network responds to stimulus

– Provided with correct response

– Modifies responding to match correct response

The Connectionist Approach

• Error signal

– Difference between actual activity of each output unit and the correct activity

The Connectionist Approach

• Back-propagation: error signal transmitted back through the circuit

• Indicates how weights should be changed to allow the output signal to match the correct signal

• The process repeats until the error signal is zero

Caption: Learning in a connectionist network. Bars represent activity in the eight representation units. Notice how the pattern of activation changes as learning progresses.

The Connectionist Approach

• Slow learning process that creates a network capable of handling a wide range of inputs

• Learning can be generalized

The Connectionist Approach

• Graceful degradation: disruption of performance occurs gradually as parts of the system are damaged

Categories in the Brain

• Different areas of the brain may be specialized to process information about different categories

– Double dissociation for categories “living things” and “nonliving things”

Categories in the Brain

• Categories are represented by distributed activity

– More similar patterns of brain activity for categories with similar features

– Category-specific neurons

Caption: (a) Testing procedure for determining if monkeys can categorize cats and dogs. (b) Activity recorded from a neuron in the monkey’s IT cortex during the testing sequence. This neuron responds more to a 60 percent dog stimulus than to a 60 percent cat stimulus during presentation of the sample. (c) Activity recorded from a neuron in the monkey’s PF cortex during the testing sequence. This neuron responds better to the dog stimulus during the delay and test periods.