Knowledge Flashcards
What does obcject recognition rely on?
(conceptual knowledge)
Bottom-Up
-> Visual Analysis of a stimulus
Top-Down
-> Knowledge structures to match the input to
What is a concept / category?
(conceptual knowledge)
Concept
-> Mental representation of a class or individual
-> used for different cognitive functions
Category
-> Collection of all possible examples of a perticular
concept
Categorization
-> The process by which things are placed into categories
Why are categories useful?
(conceptual knowledge)
=> knowledge is organized in categories
= Categories as “pointers to nowledge”
-> allows acces to the knowledge associated with the
concept of the category
-> store information common to all category members
- helps generelize new cases
- crucial for everyday life and survival
- focus on what makes individual members different
from others (efficient)
How is category membership determined?
(approach to categorization)
- Definitional approach
- Protype approach
- Exemplar approach
What is the definitional approach to categorization?
(approach to categorization)
= Determine category membership based on whether the object meets the category definition
-> does not work well
-> not all members of everyday categories have the same defining features
Definitions often do not include all members of a category
-> instead members of a category resemble one another in several ways
-> not all members will be similar in all ways, but in some
What is the prototype approach to categorization?
(approach to categorization)
= Determine category membership by comparing an
object to a prototype from the category
-> Abstract representation of the typical member of a category
-> has characteristic features describing what members of that concepts are like
Prototype as the average of commonly experienced
members of a category
-> created by adding together many pictures
-> if objects have some commonalities, they appear as a pattern in the average
What is the attribute of prototypicality?
(approach to categorization)
High prototypicality
-> Category member closely resembles the prototype
Low procotypicality
-> Category member doesn’t closely resemble the
prototype
Prototypicality is correlated with family resemblance
-> Typical objects share more properties than untypical ones
Whay are prototypical objects “special”
(approach to categorization)
Typicality effect
(Medin & Smith, 1984)
-> sentence verification technique
(compare un/typical object on category membership)
=> highly prototypical objects are judged quickly
Prototypical objects are named first in category
(Mervis, 1976)
Priming of prototypes
(Rosch, 1975)
Task: judge whether two color patches are same color
Prime: Hear the word green
Result:
-> fastest when a prototypical green was presented
-> slower for a light green
Interpretation
-> Prime activated prototypical image of green in
participants minds
What is the exemplar approach to categorization?
(approach to categorization)
= Concept is representated by multiple exemplars
-> exemplars are actual category members a person
has encountered
-> to categorize, new object is compared to stored
exemplars
What is the favored approach between Prototype and exemplar ?
(approach to categorization)
= the more similar an object is to the category prototype or a known exemplar, the faster it will be categorized
Exemplar
-> takes atypical cases into account and does not discard
information through averaging
-> represents highly variable categories better
Prototype
-> represents large categories better
=> We initially use prototype approch and later add
exemplars
How are categories hierarchically organized?
(hierarchical categorization system)
Some categories are more general than others
-> categories form a hierarchical system with at least
three levels
- global
- basic
- specific
What makes the basic level of a category special?
(hierarchical categorization system)
= optimal balance between being informative and economic when listing objects of one category
Free naming Task
-> Basic level names are preferred when naming objects
Category verification task
-> Category name first than image of object
-> Decide whether object belongs to category or not
=> fastest for basic-level category
Children
-> learn basic level concepts sooner
=> Basic level used in everyday language
=> Different cultures use same basic-level categories
Are objects categorized uniform for every one?
(hierarchical categorization system)
Expertise
-> changes preferred level of response
Factors that influence our experience
-> Age
-> Gender
-> Culture
-> Profession
-> Object availabiity
What is a semantic network
(semantic networks)
Sematic Networks
(Collins & Quillian, 1969)
Model of how the mind represents and organizes categories and concepts
Consists of
-> Nodes as Categories
-> Links between Categories
-> Properties of the categories
Network is hierarchical, related concepts are conected
Principle of cognitive economy
-> shared properties are stored at higher nodes and
inherited
-> Exceptions are stored at lower nodes
What is the principle of cognitive economy?
(semantic networks)
= Shared properties are stored at higher nodes and inherited
Example: Canary
Canary: Can sing, is yellow
Bird: Can fly, has wings, has feathers
Animal: Can move, has skin
Living thing: Can grow, is living
-> by not storing “can fly” for each bird, system “saves storage space”
What is the evidence in support of Semanic networks?
(semantic networks)
Collins & Quillian, 1969
Prediction:
-> Time to retrieve information about a concept should
be determined by the distance one would have to
travel in the network
Task
-> Sentence verification task using category membership
and feature descriptions at different levels
- canary is canary (0-level jump)
- canary is a bird (1-level jump)
- canary is an animal (2-level jump)
Results
-> Reaction Time well predicted by distance
Meyer & Schvaneledt, 1971
Prediction
-> if a node is activated, activation will spread along the connected links (spread activition)
=> Conceps that receive activation are primed and more easily accessed from memory
Task
-> Lexical decision task
(decide if both words are real english or not)
Result
-> Reaction time of words associated in same category faster than not associated words
What criticism do Collins and Quillian’s model face?
(semantic networks)
Cannot explain typicality effect
-> apple / pomegranate is a fruit not equally fast
responded although same distance from fruit
Cognitive economy and inheritance of properties called into question
-> same distance properties have different RTs
-> properties are not equally frequent
What is the parallel distributed processing model?
(connectionist network aproach)
Model inspired by neural architecture
-> circle = neuron-like units
-> lines = axon-like connections between units
Signal flow: Input to hidden to output units
-> Input units: Activated by stimulation from environment (receptor)
Connections have different connection weights
-> weights determine how strongly incoming signal will activate the next unit
Model represents concepts by distributed activity across network
-> some units are activated more strongly compared to others
How is a PDP model structured?
(connectionist network approach)
McClelland & Rogers, 2003
Representing concepts and properties
Knowledge is represented in seperate nodes
-> Concepts
-> Relations
-> Properties
Knowledge is reconstructed
-> activate the concept and the relation
-> activity spread along the network (determined by
weights)
-> activation from concept and relation units converge in
hidden layers
-> property units are activated accordingly
How does the network learn?
(connectionist network approach)
Two critical features for learning in these models
-> weights determine the activation of following units
-> sending an error signal backwards (back propagation)
Untrained model (all weights equal)
= All representation units respond equally to input
Canary -> Hidden1 -> fly, sing, swim
Can -> Hidden1 -> fly, sing, swim
(Back propagation of error signal)
Trained model (weights adjusted, many repetitions)
= Representation units respond similarly to similar input
Canary -> Hidden1 -> fly, sing
Can -> Hidden1 -> fly, sing
(When input from these two units received, dont send signal to that unit)
What are the advantages of connectionist networks?
(connectionist network approach)
Biological plausibility
-> Proposed structure and function is based on
information representation in the brain
Can explain several controversial findings
-> typicality
-> graded category membership
-> generelization of learning
Gradual disruption of performance
-> Damage to a network does not destroy it completely
=> Strong influence on cognitive science and machine learning / AI development
