Week 2 Flashcards
Endel Tulving
Episodic Memory: autobiographical knowledge about personal past, unique to the individual
Semantic Memory: general knowledge about the world that all members of a culture possess
Penguin
First saw penguin at toronto zoo at 1985 (Episodic)
A bird that cannot fly (Semantic)
Definitions
Conceptual knowledge doesn’t remember specific events like birthdays but general things such as what consists in a wedding
You have the idea what is involved in particular event or object
Allows us to use this knowledge to understand what we are looking at and what is the appropriate actions to perform.
A concept is a mental representation of an idea.
A dog is a concept
General idea, category is specific exemplar
Cateogorization is putting things into groups
Category of cats for example
Very related
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
Allows us to deal with new events or objects
Experiencing something new, got to understand what you are going to do in this situation
Use previous knowledge and react according to what the similar thing was
Figure out what the appropirtate response is
E.g., a bird has a beak, wings, creates nest, lays eggs etc
Allows us to figure out unique characteristics
Within a category, there are exceptions such as birds that cannot fly
Categories are useful for guiding us
Definitional Approach
Determine category membership based on whether the object meets the definition of the category
* Let’s make a definition for “chair”
* A chair is a piece of furniture for one person to
sit on. Chairs have a back and four legs.
Idea that what you are doing, you figure out what the rules are for that category
Essentially try to define that category
E.g., the definition of chair
Typically a chair is a piece of furniture, meant for a single person
We look at different objects and see if they fit into the chair category
Hard to properly define a category
In practice, this not work quite well
People realized that the better way of approaching this would be resemblance, things in this category are similar with eachother
Prototype Approach
Prototype - an average of category members encountered in the past
* “typical” member of a category
* Characteristic features that describe what members of that concept are like
Ellanoar rosh
We determine if something is apart of a category if our item is an average of categories from the past
E.g., ideas of birds and averaged together and create a prototype
Essentially the typical member of the category but this average does not actually exist
This is computed from mind but there is not actual prototype
Average will have features that will are mostly common with members of that category
Prototype Perception
Closer to the prototype the easier to process
* Fluent processing of a stimulus can be mistaken for:
* Recognition
* Preference
* Truth and Beauty
Idea is that the closer the item is to the prototype, the easier it is to process it into the category
E.g., two patches of green or light green, faster to process the prototype green
Typicality effect
Number of studies that looked at fluency of processing and the things that are processed most fluently, it is something that was experienced before
Larry gecobiate, gave people list of names and had to make decisions about them and the next day, they were given another list and that list contained famous names and some from the previous day along with new names and had to decide who the famous people but when looked at the new names, they were much more likely to falsely claim that the previous day people were the famous people (false fame effect)
Situations in zyomph, looking at line drawing objects such as furniture, presented these at 1 millisecond per item and replaced by new item, for the person watching, this was a blur, after watching this, ziomph asked people to decide preference of chair which one of was in the blur and people chose the one from the blur
Now that item is processed more fluently
Propaganda effect, sir francis calton in 1860s, he took faces of a bunch of people and morphed them together which made an image about people and asked people for opinion of beauty, people preferred the composite face to the individual face
Idea was that the prototype (average) was found the most beautiful
Matches the face the best hence why you chose it
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
Idea is what you store all experiences that you had placed into a specific category in memory and when a new memory is formed, you compare it to a subset of existing memories
Multiple comparison
Not comparing to the average, comparing to actual members of categories
J
udging how similar it is to every members then make decision
Cont.
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 (family resemblance effect)
Both of these approaches talk about similarity, no specific rules that must be met
Cont. 2
Easily takes into account atypical cases
easily deals with variable categories
Basically saying how well does it resemble to category
Compare it to average and each member
Both can explain typicality effect since the more similar it is, the better it is going to match and faster to process it.
Exemplar effect has advantage of taking atypical cases into account
Such as larger amounts of things, exceptional cases
You can also deal with categories that are very specific
E.g., category of games
Idea of what games are but there are many things that take into account
Very hard to create a prototype of this, but with exemplar, all these instances can be compared to the new item and see if they match
Prototype or Exemplars
May use both
Prototypes may work best for larger categories (general)
Exemplars may work best for small categories (exceptions)
Probably use prototypes initially, since its easy and might work for categories that are general
Exemplar is good for taking care of exceptions,
beneficial for smaller categories since you do not need to compute an average
Recognition & Recall
Production Vs. Verification
Verification = indicating the truth of a test item
Production = retrieving an instance from memory when given a cue
People look at reaction time since accuracy is at ceiling
You have to say if this fits: e.g., fruit - peach,
Analgouge of recognition
Analgouge of recall is production
E.g., fruit - A
People use these tasks in investing semantic memories
Allan Collins & Ross Quillian
1969, 1970
First model of semantic memory
Hierarchical Network Model (Collins & Quillian, 1969)
Semantic memory consists of a network of basic elements (nodes) connected by pointers which express relations between elements
Stored with each element is a list of properties that define the features of each concept
Claimed that semantic memory is a network, involves nodes and these nodes are connected to one another
These connections describe the relationship between elements
Argued that concepts are the nodes themself but connected to the elements are the properties associated with these elements (features apart of the concept)
Organization of these elements are in the hierarchy
General point at top and it gets more specific as you go down
Very limited in computing powers and digital, so they said that the system is trying to be sufficient as possible,
add cognitive economy to help
You store features at highest level of hierarchy.
Everything below it inherits the features
Semantic Network
e.g., Animal -> skin, moves
Bird -> wings, flies
HNM
You can enter the model at some particular hierarchy and make a decision between 2 elements, it takes time to move levels
If you want to access features at any level, that will also take time (comparing something in lower level, must go up)
Argued that the more levels you have to move, slower the responses will be
Faster to make decisions about category membership, than for properties in higher hierarchy
6 different questions for the example
Canary to bird is one up on hierarchy
Bird to animal is another one up
P2 means you have to go up 2 levels
HNM Test
Tests of the model: Sentence verification. True? Yes or No
Category Membership (supersets)
S0: a canary is a canary
S1: a canary is a bird
S2: a canary is a animal
Properties:
P0: a canary is yellow
P1: a canary can fly
P2: A canary has skin
Results (Collins 1970)
Higher on hierarchy means slower reaction time (property)
Green is lower than red so green is faster than property (superset faster than property)
Priming
Found that people who got canary can fly were faster at making decision about canary is a bird compared to people who had canary is yellow
1 deficiency in this model
Problems with HNM
Model does not explain the typicality effect: Faster to verify typical members of a category than atypical members
In model, typical and atypical members are at same level of hierarchy, so should take the same time
Not every member of category was equal
If we make people say robin vs. chicken is a bird, people will say robin
But in hierarchy model, all of them would be at lower level so none of them would preferred so they would have the same response time
More Problems
Another study found problematic results
Is a pig a animal were quicker than a pig is a mammal
1268 vs. 1476
Each level that you traverse should take extra time
No reason why two levels should be quicker than 1 level
Feature Comparison Model (Ripps, Shoben, Smith, 1973)
Developed to explain results that HN cannot explain
All concepts in SM are represeted as sets or lists of features
No cognitive economy, same features stored with different concept
FCM
Argued that there is no hierarchy
Each concept was a unique item and each node has all features/properties connected to it
Argued that each concept, there were two types of features
Defining features - things that are needed to define the concept (bird - feathers)
Characteristic features - features that are common to many members but are not essential (birds - fly)
You have to figure out if there is a relationship between concepts
There is no physical connection between concepts in this model
You have to compute if there is a relationship
You do this by comparing the features
FCM: Stage 1
Idea is you go to robin category and retrieve features from that category and then go to concept of birds and get those characteristics and then you look at the common alibies
Are the features retrieved from robin also the ones received from bird?
If match is very low, less than 20% e.g, make decision that there is low similarity and if there are high, 80%, then decide if things are highly likely related
Low = false
High = true
At second stage, only look at defining feature
Stage 2
When you have typical member of category, it shares a lot of features
E.g., robin and bird
A pencil is bird would draw very little similarities
Ostrich is a bird would draw medium similarity as it shares defined but characteristic features
Samething with bat but this time its similar with characteristic features but not actual defining features
FCM
This can easily explain typicallity effect but cannot define priming effect
There is no memory of what you have done on FCM, no reason of why what you did earlier should impact what you did now
You retrieve characteristic and defining feature of canary but not yellow
Spreading Activation Network Model (Collins & Loftus, 1975)
Concepts organized in network, but organization not hierarchical
Features also stored with concepts
Length of links between concepts represents strength of associations
Assumes activation spreads between concepts
Essentially revised hiearcheral model
Recognized there is a network of semantic information
Also abandoned cognitive economy idea, argued that we have concepts and features and all are nodes that are connected with eachother
Closer to one another in hierarchy
Idea is figuring out if two thing are related in hierarchy
Example
All are put in network
Things that are highly associated with another will be closer than things that are less related
E.g., cherries are closer to apples than red
Explain technicality effect as cherries and red would cause activation as there is a relation and meet up fairly quickly
Longer the connection, longer the time it takes to establish connection
Explain priming since activation remains in system, faster to establish length the second time
An improvement
Meyer and Schvaneveldt (1971)
Lexical Decision Task
RT measure
NO trials (usually) not of interest
Are YES trials faster when preceded by a related word? Yes = PRIMING
Task required people presented with things on screen and asked to determine if it was a word or not
Measured reaction time
Focused on the yes responses
Presented 2 items and determine if they were words or not
E.g., Fundt and Glurb. Chair, money, nurse, doctor etc
Some pairs of words were related to one another
Associated words caused faster responses
This was evidence of the priming effect
Activation in these 2 nodes and this connection is better and closer
Priming will spread and lead to priming effect quicker
Neely (1977)
Presented indivuiduals with words with specific category and another word where they had to make a lexical decision
Timing between varied (150 msec to 2000msec)
Conditions were involved
No shift - if particular prime word was given, know ahead of time that lexical word would be a member of the category
E.g., bird as prime (no shift), ⅔ is robin, other ⅙ of time, it would be other category would be body part or building piece so ⅓ of either
If prime word was body
⅔ building, ⅙ body and ⅙ bird, so if body was given, expect building (shift)
Building as prime would be ⅔ of body, ⅙ of building and ⅙ bird
These shifts varied across the individuals
How much of semantic priming is due to automatic processing and how much is due to control processes
Results
No shift - bird ⅔ -
Bird and robin, across the board, there are benefits since they are close
Faster to verify target as they were both members of the prime category
Faster at short and long SOA
Bird-arm: performance deteriorated
Results suggested that priming effect benefited the bird group
Suggested automatic spreading activations at the short SOAs
Suggested that long SOAs has combination of automatic processes and expectanceis
Violates expectations which delays reaction time since it wasnt the item that you were expecting
Initial evidence that automatic spreading occurs and as time increases, expectancy effects come into play
Shift Condition
Body - ⅔ building
For the thing you expect, at the shortest soas, no benefit and no cost
In order for benefit, you have to make connection of that body ⅔ time is building, essentially figure out the connection
With a short SOA, you don’t have enough time
Body and heart, earliest SOAs, there is a facilitation effect
Body and hand leads to a benefit even though body parts only happen ⅙ of the time
Even though it was a very rare event, there was a benefit
Over time, benefit went away and became a cost
S
hows that there was a initial activation of automatic processing but eventually expectation leads to a defit to cost
Although sparrow still had benefit
Connectionist Approach
Idea that you have network of nodes that represents how things are connected and know idea is that model is supposed to be better analouge to the brain
Spreading activation network, concepts are each of nodes
E.g., cat doesnt activate single point in brain, it activates a lot since its a distributed network
Created network that has activation associated with particular concept is represented by activity across entire network
People argue that between nodes and network, there are links and the strength of them are based on how close things are
The more times neurons are fired, the stronger the connection will be with nodes
If one fires, it decreases
Cont.
Input nodes come from outside world
Output nodes are from hidden units and determines the actual response from network
Network sends activity to connected node and depending on strength could cause them to fire which fire onto output nodes
CA
How learning occurs:
- network responses to stimulus
- provided with correct response
- error signal
Different between actual activity of each output unit and the correct activity
Modifies responding to better match correct response via back propagation
The process repeats until error signal is 0
Training Phase: give network particular stimulus, sometimes there are educated guesses but often just randomly set all the weights
Provided input on gets to hidden and output
Compare output to correct response and calculate the error signal and see how off the current settings are
Adjust connection between the nodes
Idea is that modifying the weights, you can get actual settings better so in future you can get more accurate responses at output level
Neural Networks: Error is calculated at the very end, not at each layer of the neural network
