Knowledge Chapter 9 PDF

Knowledge Chapter 9  Knowledge Concept: Your knowledge about a particular thing or event;; a unit of knowledge g in semantic memory. y □ A chair has a seat, back, 4 legs and is somethin...

Knowledge Chapter 9  Knowledge Concept: Your knowledge about a particular thing or event;; a unit of knowledge g in semantic memory. y □ A chair has a seat, back, 4 legs and is something you sit on.  Knowledge Concept: Your knowledge about a particular thing or event;; a unit of knowledge g in semantic memory. y Category: A group of related concepts in semantic memory. □ Dining room chairs and living room chairs are both part of your chair category. Categorization: The process of assigning a new piece of information to one of these groups.  Knowledge What are categories good for? Categories help us to identify objects objects. □ They confine recognition to a smaller group of items, thereby making it faster and more accurate.  Knowledge What are categories good for? Categories help us to identify objects objects. Categories allow us to ignore the variability between the objects in a group. □ Our category for the letter “A” captures the essential features of that pattern, allowing us to ignore the non-essential variability. variability  Knowledge Color Variability Our visual system is capable of discriminating about 7 million different colors, yet we only use about 7-10 color names. Our color categories allow us to ignore subtle differences in color.  Knowledge What are categories good for? Categories help us to identify objects objects. Categories allow us to ignore the variability between the objects in a group. Categories reduce the need for constant learning. □ We don’t need to be explicitly taught about every object in the world because we can recognize that object based on its similarity to an existing category. □ Categorization g frees us from the need to encode the detailed features and properties of each new object, we simply encode that object as another member of one of our g categories.  Knowledge What are categories good for? Categories help us to identify objects objects. Categories allow us to ignore the variability between the objects in a group. Categories reduce the need for constant learning. But categorization is also responsible for a lot of memory errors. □ By categorizing an animal as a sort of squirrel, it inherits all of the “squirrel” squirrel properties. properties You may therefore remember the animal as having a bushy tail, even if it did not.  Knowledge How is categorization studied? Definitional Approach: Define minimal criteria that an object must have to be included in a category. □ A “plane” has to have an engine and wings and be able to fly. □ But what are the defining criteria of a chair?  Knowledge How is categorization studied? Definitional Approach: Define minimal criteria that an object must have to be included in a category. □ A “plane” has to have an engine and wings and be able to fly. □ But what are the defining criteria of a chair?  Knowledge How is categorization studied? Definitional Approach: Define minimal criteria that an object must have to be included in a category. Family Resemblance: Members of a category are similar to each other in a large number of ways, but any one way is not usually essential. □ C Categorization t i ti iis b based d on th the similarity i il it b between t th the new object and the members of each existing category. □ Some found this implausible given the many comparisons that would be needed to compute family resemblance.  Knowledge How is categorization studied? Definitional Approach: Define minimal criteria that an object must have to be included in a category. Family Resemblance: Members of a category are similar to each other in a large number of ways, but any one way is not usually essential. Prototype Approach: New objects are compared to each category’s prototype; objects are classified based on the best match.  Knowledge Rosch (1975) A prototype is the “average” average of a category category’s s membership. membership □ What you consider to be a “good bird” based on the members of your “bird” category. categor members category prototype  Knowledge Rosch (1975) A prototype is the “average” average of a category category’s s membership. membership But the prototype is rarely an actual member of a category. categor members category prototype  Knowledge Rosch (1975) A prototype is the “average” average of a category category’s s membership. membership But the prototype is rarely an actual member of a category. 86 64 92 The average is set of exam 96 not a member of scores 78 this set. 60 82 mean 79.7  Knowledge Rosch (1975) A prototype is the “average” average of a category category’s s membership. membership But the prototype is rarely an actual member of a category. The prototype constantly changes with each new exemplar encountered or object added to the category. □ Just as an arithmetic mean changes with each new item added to a set. 98 some 62 62 exam 90 90 scores 86 86 mean 79.3 84.0  Knowledge Rosch (1975) A prototype is the “average” average of a category category’s s membership. membership But the prototype is rarely an actual member of a category. The prototype constantly changes with each new exemplar encountered or object added to the category. Some new objects will match the prototype well (high prototypicality), others will not (low prototypicality).  Knowledge Rosch (1975) category members prototype high prototypicality new object is likely to object be categorized as a bird bi d  Knowledge Rosch (1975) category members prototype low prototypicality new object is less likely object to be categorized as a bird bi d  Knowledge Rosch (1975) A prototype is the “average” average of a category category’s s membership. membership But the prototype is rarely an actual member of a category. The prototype constantly changes with each new exemplar encountered or object added to the category. Some new objects will match the prototype well (high prototypicality), others will not (low prototypicality).  Knowledge Techniques for Studying Prototypicality Object Naming: Subjects are asked to name members of a given category (“birds”). Typical members (“robin”) are named before less typical members (“penguins”). Prototype Priming (Rosch, 1975): Primed fast subjects with a color name (“green”), then asked them to respond slow whether hether two t o colors were ere the same or different. RTs were faster when the colors matched the prime prototype.  Knowledge Techniques for Studying Prototypicality Object Naming: Subjects are asked to name members of a given category (“birds”). Typical members (“robin”) are named before less typical members (“penguins”). Prototype Priming (Rosch, 1975): Primed subjects with a color name (“green”), then asked them to respond whether two colors were the same or different different. RTs were faster when the colors matched the prime prototype. Feature Overlap Analysis: Subjects list attributes for several objects under a category, then the experimenter determines which objects have attributes in common; more attributes in common the greater the typicality. □ “Cars” have lots of attributes in common with other members of the vehicle category, but “elevators” do not (atypical).  Knowledge Techniques for Studying Prototypicality Category Verification Task (Smith et alal., 1974): Subjects would see a category name (“fruit”), then shown a picture of a object (“apple”). The task would be t indicate, to i di t as quickly i kl as possible, ibl whether h th th the object bj t was a member of the category. FRUIT → → “yes” FRUIT → → “no” “ ”  Knowledge Techniques for Studying Prototypicality Category Verification Task (Smith et alal., 1974): Subjects would see a category name (“fruit”), then shown a picture of a object (“apple”). The task would be t indicate, to i di t as ffastt as possible, ibl whether h th th the object bj t was a member of the category. RTs were faster to objects rated as highly typical of the category compared to objects rated as less typical.  Knowledge Hierarchical Organization of Categories (Rosch et al. 1976) Superordinate (global) Level: A very broad level of categorization (furniture, vehicles, musical instruments). Basic Level: A middle level of categorization under the superordinate (chairs, tables, beds). Subordinate S b di t (specific) ( ifi ) Level: L l Detailed D t il d categories t i under each basic (dining room chairs, office chairs).  Knowledge The Basic Level Advantage (Rosch et al., 1976) Categorization starts at the basic level in this hierarchy hierarchy. Used a category verification task and varied whether the category name was superordinate, basic, or subordinate. superordinate verification FURNITURE → → “yes”  Knowledge The Basic Level Advantage (Rosch et al., 1976) Categorization starts at the basic level in this hierarchy hierarchy. Used a category verification task and varied whether the category name was superordinate, basic, or subordinate. basic verification CHAIR → → “yes”  Knowledge The Basic Level Advantage (Rosch et al., 1976) Categorization starts at the basic level in this hierarchy hierarchy. Used a category verification task and varied whether the category name was superordinate, basic, or subordinate. subordinate verification DINING ROOM CHAIR → → “yes”  Knowledge The Basic Level Advantage (Rosch et al., 1976) Categorization starts at the basic level in this hierarchy hierarchy. Used a category verification task and varied whether the category name was superordinate, basic, or subordinate. mismatch condition CHAIR → → “no”  Knowledge The Basic Level Advantage (Rosch et al., 1976) Categorization starts at the basic level in this hierarchy hierarchy. Used a category verification task and varied whether the category name was superordinate, basic, or subordinate. Verification times were fastest at the basic level. Basic Level Superiority Effect FURNITURE → → “yes” yes slow s o CHAIR → → “yes” fast DINING ROOM CHAIR → → “yes” slow  Knowledge The Basic Level Advantage (Rosch et al., 1976) Categorization starts at the basic level in this hierarchy hierarchy. Used a category verification task and varied whether the category name was superordinate, basic, or subordinate. Verification times were fastest at the basic level. Using a naming task, Rosch found that subjects name objects at the basic level. EXPERIMENTER SUBJECT “What is this?” → → “chair”  Knowledge The Basic Level Advantage (Rosch et al., 1976) Categorization starts at the basic level in this hierarchy hierarchy. Used a category verification task and varied whether the category name was superordinate, basic, or subordinate. Verification times were fastest at the basic level. Using a naming task, Rosch found that subjects name objects at the basic level. Rosch also found that babies first start to say words at the basic level level.  Knowledge Tanaka & Taylor (1991) Asked whether experts also show a basic level advantage advantage. Had bird experts and non-experts do an object naming task for a variety of categories, including birds. Bird experts named birds at the subordinate level (robin), whereas non-experts used the basic level name (bird). BIRD EXPERT EXPERIMENTER → “hummingbird” hummingbird “What is this?” → → “bird” NON-EXPERT  Knowledge Tanaka & Taylor (1991) Asked whether experts also show a basic level advantage advantage. Had bird experts and non-experts do an object naming task for a variety of categories, including birds. Bird experts named birds at the subordinate level (robin) (robin), whereas non-experts used the basic level name (bird).  Knowledge Tanaka & Taylor (1991) Asked whether experts also show a basic level advantage advantage. Had bird experts and non-experts do an object naming task for a variety of categories, including birds. Bird experts named birds at the subordinate level (robin), whereas non-experts used the basic level name (bird). Experts organize information so as to enable preferential access to their domain of expertise. □ Their basic level categories are what we consider subordinates, and their subordinate level categories are things we probably never heard of.  Knowledge What causes the basic level advantage? A “good” good category should have two things (Rosch et al al, 1976): □ The members of a category should share lots of attributes with each other (everything in your dog category should have 4 llegs and d some ki kind d off b bark; k th they shouldn’t h ld ’t meow or chirp). hi ) □ The members of one category should not share attributes with the members of a different category (members of your dog category should not fly or swim underwater).  Knowledge What causes the basic level advantage? The superordinate level is good at satisfying the “minimal minimal overlap” criterion (members of different categories share few attributes); animals don’t overlap with furniture. The superordinate level is bad in that its members don’t have many attributes in common (what do all members of the furniture category have in common?) common?). The subordinate level is good in that its members have lots of attributes in common (all wing chairs are chairs having two wings on the upper back). The subordinate level is bad at satisfying the minimal overlap criterion; wing chairs share a lot of attributes with dining room chairs, office chairs, etc.  Knowledge What causes the basic level advantage? A “good” good category should have two things (Rosch et al al, 1976): □ The members of a category should share lots of attributes with each other (everything in your dog category should have 4 llegs and d some ki kind d off b bark; k th they shouldn’t h ld ’t meow or chirp). hi ) □ The members of one category should not share attributes with the members of a different category (members of your dog category should not fly or swim underwater). The basic level finds a happy medium between these two criteria. criteria □ Most members of the basic “chair” category have many attributes in common (a seat, back, place to rest arms, etc.). □ But chairs have minimal overlap with other categories at the basic level (chairs are different than lamps, tables, and beds).  Knowledge Rosch et al. (1976) Had subjects list common attributes for categories at the superordinate, basic, and subordinate levels. Subjects could think of only a few attributes (0-3) that several superordinate categories had in common. But subjects were able to list a lot of common attributes at the basic level (average of 9) 9). Subjects list more common attributes at the subordinate level, but not many; a small benefit is more than offset by increased attribute overlap.  Knowledge Rosch et al. (1976) Had subjects list common attributes for categories at the superordinate, basic, and subordinate levels. Subjects could think of only a few attributes (0-3) that several superordinate categories had in common. But subjects were able to list a lot of common attributes at the basic level (average of 9) 9). Subjects list more common attributes at the subordinate level, but not many; a small benefit is more than offset by increased attribute overlap.  Knowledge The Hierarchical Model (Collins & Quillian, 1969) How is information retrieved from a semantic network?  Knowledge The Hierarchical Model (Collins & Quillian, 1969) Information is organized into a hierarchy of concepts concepts, just as we discussed for categorization. At each category there is a list of associated attributes. “Nodes” Nodes (categories) are connected by “links” (arrows). Attached to each node are “properties” (attributes).  Knowledge The Hierarchical Model (Collins & Quillian, 1969) The properties attached to a node apply to that node and every linked node under it in the hierarchy (fish swim, so we know that sharks can swim because they are fish). Cognitive economy: Properties are only represented once (not repeated at each node); lower nodes i h it properties inherit ti ffrom higher nodes.  Knowledge The Hierarchical Model (Collins & Quillian, 1969) The properties attached to lower nodes do not necessarily apply to linked nodes higher in the hierarchy (although birds have wings and can fly, not all animals can fly). A hierarchical organization, combined with the principle of cognitive economy is a very efficient ffi i t method th d off representing information.  Knowledge The Hierarchical Model (Collins & Quillian, 1969) Efficiency comes with a price; because information is distributed throughout the hierarchy, the hierarchy must be navigated when retrieving information.  Knowledge The Hierarchical Model (Collins & Quillian, 1969) Efficiency comes with a price; because information is distributed throughout the hierarchy, the hierarchy must be navigated when retrieving information. Navigation also applies to properties. To verify the truth of: “canaries can fly”, you would need to navigate i t tot the th bird bi d node were “can fly” is located.  Knowledge Assumptions of the Hierarchical Model The model assumes that in order to retrieve information you must be at the node in which the information resides. Also, if a property listed at a higher node is not true for a linked lower node node, then an exception must be listed at the lower node (“can’t fly” is listed at the “ostrich” node).  Knowledge Assumptions of the Hierarchical Model Retrieving a piece of information (such as the property “is yellow”) takes a certain amount of time. It also takes time to move from one node to another via a link ((“canary” canary to “bird”) bird ). Information retrieval times and movement times are additive.  Knowledge Testing the Hierarchical Model Which proposition should take longer to verify: “a canary can sing” or “a canary can fly”? It should take longer to verify “a canary can fly” because this requires q both a movement and a retrieval operation.  Knowledge Testing the Hierarchical Model Which proposition should take longer to verify: “a canary can fly” or “a canary is a bird”? It should take longer to verify “a canary can fly” because this requires q both a movement and a retrieval operation; verifying “a canary is a bird” just requires a movement.  Knowledge Testing the Hierarchical Model Collins and Quillian used a sentence verification task; subjects had to verify whether a proposition was true or false. There were six types of propositions differing in th number the b off retrieval ti l and d movementt operations. ti EXPERIMENTER SUBJECT “a canary is a bird” → “true” “ canary iis a fish” “a fi h” → “false” “f l ”  Knowledge Testing the Hierarchical Model Collins and Quillian used a sentence verification task; subjects had to verify whether a proposition was true or false. There were six types of propositions differing in th number the b off retrieval ti l and d movementt operations. ti 1 property retrieval operation (“can sing”); 0 movement operations. 1 property retrieval operation (“can fly”); 1 movement operation (canary to bird).  Knowledge Testing the Hierarchical Model Collins and Quillian used a sentence verification task; subjects had to verify whether a proposition was true or false. There were six types of propositions differing in th number the b off retrieval ti l and d movementt operations. ti A property must be retrieved for verification. Verification does not require property retrieval.  Knowledge Testing the Hierarchical Model RTs increase with the number of levels moved moved. RTs also increase when a property retrieval operation is needed. These movement and property retrieval times add linearly (resulting in parallel slopes).  Knowledge Spreading Activation Model (Collins & Loftus, 1975) The earlier Collians and Quillian model didn’tdidn t account for typicality effects; it predicts that “a canary is a bird” would be verified as fast as “an ostrich is a bird”, when in f t “ostrich” fact “ t i h” is i a less l typical t i l bibird d and d ttakes k llonger. Also, some verification results contradicted predictions; “a pig is a mammal” takes longer to verify if th than “a “ pig i iis an animal”.  Knowledge Spreading Activation Model (Collins & Loftus, 1975) Similarities to the Collians and Quillian model: □ Both models assume that processing passes along links between concepts (canary to bird), and that this takes time. Differences from the Collians and Quillian model: □ The spreading activation model is not hierarchical. □ Properties can now be represented at multiple places in the knowledge structure; the principle of cognitive economy is not strictly enforced. □ There are no longer any property lists attached to concepts; properties (“can sing”, “is yellow”) are treated like concepts, just like “canary” and “bird”.  Knowledge Spreading Activation Model (Collins & Loftus, 1975) Once a concept is activated activated, this activation spreads to all linked concepts.  Knowledge Spreading Activation Model (Collins & Loftus, 1975) Once a concept is activated activated, this activation spreads to all linked concepts.  Knowledge Spreading Activation Model (Collins & Loftus, 1975) Once a concept is activated activated, this activation spreads to all linked concepts.  Knowledge Spreading Activation Model (Collins & Loftus, 1975) Once a concept is activated activated, this activation spreads to all linked concepts. Short links represent stronger connections than longer links. The more links attached to a concept, the smaller the amount of activation spreading from that concept down each link.  Knowledge Spreading Activation Model (Collins & Loftus, 1975) Once a concept is activated activated, this activation spreads to all linked concepts. Short links represent stronger connections than longer links. The more links attached to a concept, the smaller the amount of activation spreading from that concept down each link.  Knowledge Spreading Activation Model (Collins & Loftus, 1975) Once a concept is activated activated, this activation spreads to all linked concepts. Short links represent stronger connections than longer links. The more links attached to a concept, the smaller the amount of activation spreading from that concept down each link.  Knowledge Spreading Activation Model (Collins & Loftus, 1975) Final Assumptions: □ Activations disappears from the system over time. □ If spreading g activation results in one concept becoming g active above some threshold, you will “think” of that concept. □ Activation will then flow from that concept to all other linked concepts, causing the process to continue and our thoughts to flow from one thing to another.  Knowledge Priming and Spreading Activation Priming: When exposure to some object or event improves processing of some later object or event. Two types: □ Repetition Priming: When processing something a second time benefits from having processed it previously. Previous exposure to the word “METAL” might result in faster identification of the word fragment…  Knowledge Priming and Spreading Activation Priming: When exposure to some object or event improves processing of some later object or event. Two types: □ Repetition Priming: When processing something a second time benefits from having processed it previously. □ Associative Priming: When processing something benefits from having processed something related previously. □ Key difference: In repetition priming an item is processed twice, in associative priming a specific item is processed only once. Assumption: The more associated two things are in memory, the more those things should prime each other.  Knowledge Priming and Spreading Activation Lexical Decision Task: Subject has to report whether a string of letters is a valid English word. Experimenter presents: Subject responds: BUTTER yes BUFLER no  Knowledge Meyer & Schvaneveld (1971) Used a modified lexical decision task; subjects saw two letter strings and had to indicate whether both were valid words. Manipulated whether the words were associated or not. not associated associated  Knowledge Meyer & Schvaneveld (1971) Used a modified lexical decision task; subjects saw two letter strings and had to indicate whether both were valid words. Manipulated whether the words were associated or not. RTs were faster for the associated words (bread/wheat) compared to the unassociated nassociated words (chair/money). Interpreted as direct evidence for associative priming, and indirect evidence for spreading activation; activation spreads between the associated words, making the verification of each easier.  Knowledge Neural Network Models Spreading activation underlies currently popular neural network models of knowledge representation. Concepts are represented by patterns of activity across nodes. Associations are captured by how activity spreads from one concept to another. another