Chapter 8: Concepts and Generic Knowledge

How is the concept (knowledge) network organized in the brain? *{CLASS}*

Are concepts organized according to categories (e.g., living, non-living) or properties (e.g., visual, motor) certain types of temporal lobe damage can produce seemingly bizarre deficits in one's ability to categorize e.g., living vs nonliving -patients do fine at categorizing non-living objects (e.g., briefcase) -perform poorly at categorizing living objects (e.g, spider) =category based organization? an alternate hypothesis for this pattern is that visual features (e.g., what does a leopard look like) predominant role in our knowledge of living things (more so than for non-living) non-living objects tend to be categorized more along functional attributes (e.g., what is a desk used for?) greater damage to visual semantic memory has a disproportionate impact on categorization of living things this fits the views of connectionist architecture verbal and visual representations only linked through semantic memory semantic memory divided according to properties (.... Information about salient properties of an object, such as what it looks like, how it moves, and how it is used, is stored in sensory and motor systems activate when that information was acquired evidence from brain imaging in healthy adults -naming pictures of tools -naming pictures of animals as a result, object concepts belonging to different categories like animals and tools are represented in partially distinct, sensory- and motor property-based neural networks however, some property-based regions seem to show a categorical organization, thus providing evidence consistent with category-based formulation as well Together, the brain evidence suggests that object categories may not be entirely explicitly represented, but rather emerge from weighted activity within property-based brain regions (consistent with the connectionist view)

When we ask people to think about a category, they are in fact thinking about the prototype of that category *{CLASS}*

Make up a sentence about birds -I like to feed birds in the park -I like to feed robins in the park ((("robins" is the prototype))) -I like to feed penguins in the park (((hmmm...)))

What are concepts (2) *{CLASS}*

Maybe concepts are definitions? -e.g., dog -definition: a mammal with four legs that barks and wags its tail -but what about exceptions: dog that does not bark or that lost a limb -other concepts may be even more difficult to define

Basic-level categories *{CLASS}*

neither too general nor too specific; what we typically use; has the most utility -single word -the default for naming an object -easy to explain feature commonalities -basic categories are learned first (used by children to describe most objects)

heuristic strategy

one that gives up the guarantee of accuracy in order to gain some efficiency

picture-identification task

*{CLASS}* How do we test the prototype notion and demonstrate typicality effects? The picture-identification task: Does this picture show you a typical bird? T/F (respond "true" to robin faster than penguin)

rating tasks

-Ps are given instructions to rate how "birdy" or "doggy" a word is from a list of birds and dogs

sentence verification task

-Ps are presented with a succession of sentences; their job is to indicate (by pressing the appropriate button) whether each sentence is true or false -in most experiments, what we are interested in is how quickly Ps can do this task -e.g., a penguin is a bird vs a robin is a bird *{CLASS}* How do we test the prototype notion and demonstrate typicality effects? The sentence-verification task: True or false? Robins are birds Penguins are birds -When there is a lot of similarity between the test case and the prototype, decision can be made quickly; judgments about items more distant from the prototype take more time

The sentence verification task provides evidence that the knowledge representation is a network *{CLASS}*

-Ps must quickly decide whether a given sentence is true or false -robins are birds -robins are animals -cats have hearts -cats are birds animals - have hearts - eat food cats - have claws - purr -"cats have claws" requires one link -"cats have hearts" requires two links -reaction times are longer for judgments linking longer associative paths -properties that can link to a superordinate category may take longer when making judgments about categories at the subordinate level

family resemblance

-Wittgenstein proposed that members of a category have a family resemblance to each other -think about resemblance patterns in an actual family. there are probably no "defining features" for your family-- features that every family member has. nonetheless, there are features that are common in the family

production task

-ask Ps to name as many birds or dogs as they can -according to prototype view, they will do this task by first locating their bird or dog prototype in memory and then asking themselves what resembles this prototype *{CLASS}* How do we test the prototype notion and demonstrate typicality effects? The production task: Name as many fruits as possible Name as many birds as possible -will typically start with category members closest to the prototype (e.g., apple -- orange --- eventually tomato)

basic-level categorization

-basic level categories are usually represented in our language via a single word, while more specific categories are identified only via a phrase

What are concepts? *{CLASS}*

-concepts (categories) like colors or dogs or chairs -building blocks -simple but complex to explain -categories summarize attribute of their members Why do we care? -real world categories do not have to share identical attributes -real world categories often have continuous dimensions (colors range along a continuum for example) rather than discrete dimensions -real world categories are hierarchically organized (i.e., larger categories contain smaller categories; e.g., furniture, chair, kitchen chair) -concept identification assumes that all members of a category are equally good members. Real world categories have examples that differ in their typicality (measure of how well a category member represents that category)

local representations

-each node represents one idea, so that when that node is activated, you're thinking about that idea, and when you're thinking about that idea, that node is activated

There are different types of categories *{CLASS}*

-just as certain category members seem to be privileged, so are certain types of category -hierarchical membership; have relation to things above and below them -for example, what is this object? (pic of chair) -Furniture (detail may be too general) -A wooden desk chair (detail may be too specific) -Chair (just right) Superordinate -large category at the top of the hierarchy (e.g, furniture, tools, vehicles) -members have few common attributes Basic-level -intermediate category (e.g., table, saw, truck) -just right! -members (e.g., chair) share attributes, but also have attributes that differ from those of items in other basic-level categories (e.g., lamp) Subordinate -small category at bottom of hierarchy (e.g., throne, jigsaw puzzle, kawasaki motorcycle) -members share many attributes with members of similar subordinate categories, but very specific

Knowledge is a network *{CLASS}*

-knowledge is not randomly organized -knowledge is represented via a vast network of connections and associations between all of the information you know -related bits of information are linked together to form this network, knowledge is a product of the interactions between these bits of information -because of this organization, less related bits are further away in the network and more related bits are closer to each other -common or redundant information represented at superordinate levels -e.g., shark is closer to salmon than it is to canary -common or redundant information represented at superordinate levels animal - has skin - can move - eats - breathes bird - has wings - can fly - has feathers canary - can sing - is yellow ostrich - has long thin legs - is tall - can't fly

What are the nodes and links of the knowledge network? *{CLASS}*

-nodes can represent concepts -links such as hasa or isa can associate each concept -OR we can think about a much larger set of relationships than just equivalence ("isa") and possession ("hasa")

graded membership

-objects closer to the prototype are "better" members of the category than objects farther from the prototype -e.g., some dogs are "doggier" than others, some books are "bookier" than others

definitions

-one way to think about definitions is that they set the "boundaries" for a category -if a test case has certain attributes, then it's "inside" the boundaries -if a test case doesn't have the defining attributes, then it's outside the category

prototype theory

-perhaps the best way to identify a category, to characterize a concept, is to specify the "center" of the category, rather than the boundaries (like a definition) *{CLASS}* categorizing the category by a prototype

Prototypes and typicality effects *{CLASS}*

-prototype: one that possesses all the characteristic features of a category -can be ideal -can be average of various category members that have been encountered -differ across individuals -may differ across countries (e.g., the prototypical house in the US compared to Japan) -graded membership: some members are closer to the prototype -fuzzy boundaries: no clear dividing line for membership

connectionist networks

-rely on distributed representations, so that any particular idea is represented only by a pattern of activation across the network

propositions

-the smallest units of knowledge that can be either true or false

connection weights

-the strength of the individual connections among nodes

parallel distributed processing (PDP)

-we know that the brain relies on parallel processing, with ongoing activity in many regions simultaneously -we also know that the brain uses a "divide and conquer" strategy, with complex tasks broken down into small components, and with separate brain areas working on each component -in addition, PDP models are remarkably powerful: according to many researchers, computers relying on this sort of processing have learned the rules of english grammar, have learned how to read, and have even learned how to play strategic games -PDP models have an excellent capacity for detecting patterns in the input they receive, despite a range of variations in how the pattern is implemented -thus, the model can recognize a variety of different sentences as all having the same structure, and a variety of game positions as all inviting the same next move -as a result, these models are impressively able to generalize what they have "learned" to new, never-before-seen, variations on the pattern

Members of a given category have family resemblance to one another *{CLASS}*

Probabilistic Phrasing -there are probably no "defining" features for your family, but there are probably features that are common to your family -for common features, identify of the common features depends on the specific subgroup you're considering -when considering all features that are common across all subgroups, may arrive at an ideal, but it is not necessary that any specific individual is the ideal -a dog probably has four legs, probably barks, and probably wags its tail -a creature without these features is unlikely to be a dog -there may be no features that are shared by all dogs, just as there are no features shared by every member of a family -the more characteristic features an object has, the more likely we are to believe it is part of the category -family resemblance is a matter of degree, not all or none!

Connectionist networks use distributed representations, where information is represented by a pattern of activation across the network *{CLASS}*

Propositional networks -localist representations (each node is equivalent to one proposition) Connectionist networks -distributed processing (information involves a pattern of activation) parallel processing of information occurs at the same time learning occurs through changes in the strength of connections (the connection weights) between nodes

Conceptual knowledge uses both prototypes and exemplars *{CLASS}*

Prototypes: economical but less flexible Exemplars: more flexible but less economical -chinese vs american birds Prototypical bird: robin Exemplar bird: Big Bird Every concept is a mix of exemplar and prototype -early learning involves exemplars (when children are learning what a dog is, they think about their own dog; takes them time to realize all dogs are dogs) -experience involves averaging exemplars to get prototypes -with more experience, we can use both -the mix of using prototypes and exemplars may vary from person to person, concept to concept (maybe you have lots of experience with different fruits of the world vs someone else; you're idea of what a fruit is will differ from someone else)

Dissociations between semantic memory and episodic memory *{CLASS}*

Semantic Dementia -progressive neurodegenerative disorder characterized by loss of semantic memory in both the verbal and non-verbal domains -deficits: anomia, impaired comprehension of word meaning, associative visual agnosia, deficits in generating exemplars from categories -autobiographical memory, and day-to-day episodic memory is fine -typically associated with damage to the lateral temporal lobes (and frontal lobes) -distinct from impairments of episodic memory (amnesias), which are typically associated with damage to the medial temporal lobes -video of woman; knife, pen, pipe, scissors -knew what she was talking about, but couldn't find the words -where do you go to get your hair done? shopping center (we would say hairdresser)

Both prototype theory and the exemplar view can explain the typicality and graded-membership effects that we have discussed *{CLASS}*

Theory: Typicality Prototype: Average of a category Exemplar: Encountered more often (e.g., if you've only seen a robin once in your life, but you see Big Bird on TV all the time) Theory: Graded Membership Prototype: Less similar to average Exemplar: How often is it encountered

Are there alternatives to prototype-based categorization? *{CLASS}*

What if we made our decision by comparing the test case to a particular exemplar we have stored in memory? Does this picture show you a typical bird? -real bird photo vs Big Bird photo -your judgments are anchored in memory/things that come to mind (maybe you were watching Sesame Street earlier that day) Exemplar-based reasoning -drawing on knowledge of specific category members rather than on more general, prototypical information about the category -in both views (prototype, and exemplar), comparing to-be-categorized items to a standard. Difference between the views lies in what the standard is -for exemplar theory, standard is provided by whatever example of the category comes to mind (i.e., process of triggering memories) -different examples may come to mind on different occasions

What are concepts (3)? *{CLASS}*

concepts (categories) -building blocks -simple but complex to explain -most concepts are mixes of prototypes and exemplars concepts are hierarchically organized -superordinate -basic level -subordinate concepts represent our knowledge! but how is this knowledge actually represented?

Some simple concepts may have no precise definition *{CLASS}*

e.g., games Definition: -played by children -engaged in for fun -has rules -involves multiple people -is competitive -is played during leisure Exception: -gambling -professional sports -playing with legos -solitaire -flying kites -flying simulators

Anderson ACT network is designed around propositions *{CLASS}*

propositions: the smallest unit that can be true or false propositional networks: localist representations--each node is equivalent to one proposition -linking agents and objects and how they're related to one another dog - eat - meat dog - chase - cat can also factor in space and time

Typicality also influences judgments of attractiveness! {*CLASS*}

top photo of fish is closer to a prototype than the weird/scary fish; top photo is more attractive