Knowledge Flashcards
knowledge
scientific research is looking to emulate, understand, and duplicate knowledge
- can mimic our brain knowledge in robots
ChatGPT
knowledge as a robot
- learns writing, implicitly learns grammar
- can mimic semantic knowledge
- large language model
- can simulate knowledge or informaiton about the world
knowledge - going forward
- looking forward, scientists are working on building computers with brain cells
- the basis for the computation of information is thought to be superior if comupters can mimic the brain
to improve computer processing, need to understand:
* how we understand the world
* how to structure information
– use human knowledge as a way to understand the organization of knowledge
categories and concepts
concepts:
* mental representation of an object, event, or pattern
* decreases the amount of information we need to learn
* allow us to make predictions
category:
* class of things that share a similarity
* what are you matching incoming info to and how do you decide what’s what
theories of categorization
- definitional approach
- probabilistic views: prototype view and examplar view
definitional approach: forming concepts
form concepts by finding necessary and sufficient features
* these are defining features
what makes a square a square?
* something that doesn’t meet the definitions wouldnt fit in the category
rigid way about forming categories
* all or none (either in the category or not)
* good starting point but doesn’t reflect behavior
definitional approach: creating categories
categories have rigid boundaries
* either is a square or isn’t
all members are equally good examples
* learning involves discovering defining features
definitional approach
- people do create and use categories based on a system of defining features and rules
- many categories do not seem to follow this process
problems with definitional approach
difficulty coming up with defining features
* wittgenstein (1953): what is a game?
not all members are equally good examples
disagree on members of categories
* how would you categorize bookends?
probabilistic theories - prototype theories
category decisions made based on an idealized average – a prototpe
prototypes: have an ideal of what each category is made of
probabilitic theories - exemplar theories
an alternative to prototype theory
category decisions are made based on all of the exemplars (examples) stored in semantic memory
exemplar: categories are based on all the examples you have in your head
prototype view
- idealized representation
take all dogs and make a prototype for all of them together
the prototype approach
Bird
high-prototypicality: robin
low-prototypicality: penguin
typicality effects
Typical:
– is a robin a bird
– is dog a mammal
– is diamond a precious stone
Atypical
– is ostrich a bird
– is whale a mammal
– is turquoise is precious stone
Atypicals have slower processing times bc they are further from the prototype
graded structure
categories and concepts do not have clear all or nothing boundaries. some members of a category are more central and some are more peripheral to the prototype.
typical items are similar to a prototype
typicality effects are naturally predicted
prototype lives at center
– measure prototypicality of an object by looking at the distance of any one object to the prototype
** can quantify typicality within a category
problems with prototypes
prototypes organized around averages
information about individual exemplars is lost
* eg. Pomeranian more similar to cat than great dane
when all you store is the average, you lose nuances
exemplar view
- store all the instances (or exemplars) of category
- prototype is generated/abstracted as needed (not stored)
– stores all experiences, not just prototype
– prototype is generated as needed
– could have a prototype for dog and also for specific types of dogs
the exemplar approach
- explains typicality effect
- easy takes into account atypical cases
- easily deals with variable categories
characteristics of categories
graded membership: robin is a better bird than penguin
family resemblance: category members typically share a set of common features
related concepts:
* central tendency (prototype)
* typicality effects
organization
show pic of bird and ask what it is
people say one of:
* bird (most common)
* parrot
* grey parrot
* animal
levels of categorization (Rosch)
subordinate level: most specific level
Basic: mid level
superordinate: most broad level
basic level
the level the members share the most of the attributes of that category
“Bird” is a basic level for the picutre – most common answer
faster at processing at the basic level
subordinate level
more specific than the basic level response
“parrot” is a subordinate level
superordinate level
broad and more general than the basic level
“animal”
changes in level with expertise
experts shift from using basic level to more subordinate levels
* dog shows
knowledge influences categorization
the subordinate level becomes their fastest response
DRM Paradigm
show list of words centered around a common theme word
* people tend to insert the common theme word
collins and Quillian: Semantic hierarchical theory
semantic (memory for facts)
semantic network: information that is related is linked together
semantic network theory:
– we have nodes that represent different concepts or units of knowledge
– between the nodes is how these things are connected
- goes from more general things to more specific units of knowledge
- proposition is that all knowledge is organized in hierarchical structure
semantic hierarchical theory
nodes and links for the hierarchy
* nodes = representations of concepts
* links = representations of relationships
activation of one node spreads to related nodes
* robin would lead to activation of other nodes directly connected
* these get an activation boost
* called spreaded activation
* make the related nodes easier to process
spreading activation: excitation spreads along the connections of nodes in the semantic network
priming: primes (facilitates the activation of) related concepts — when you think about cow, its easier to process milk
spreading activation and priming
mechanisms through priming is explained through spreaded activation – lets you think of everything related
word cat activates the word dog. spreading activation activates word “dog” and primes a bit
when asked to recognize “dog” you are faster because already partially activated
sentence verification task
Collins and Quillian (1969)
* A _____ is/has a _____ statements
sentence verification task:
* a canary is an animal
* a canary is a chicken
* a canary has a yellow color
how quickly you cna verify a category depends on how close the information is
properties are assigned different nodes
* Canary: sing, yellow, and also has properties of what’s above it like bird (fly, feathers)
Collins and Quillians results
RT was longer when number of associative links was greater
less relevant property relationships, longer RT
(a canary can breathe is longer than a canary is yellow)
property inheritance: greater distance and more connections — the longer is takes
generalizations
DRM paradigm
lexical decision task
** how does spread of activation explain DRM paradigm —> boosts activation
criticisms of collins and Quillian
theory cannot explain typicality effects
* Canary is a bird
* ostrich is a bird
cognitive economy now always true
* a horse is a mammal (longer RT)
* a horse is an animal (shorter RT)
^ those go against the premise of the hierarchy
due to these criticisms, other network models were then proposed
* connectionist models
* representations of concepts
connectionist models
explains how the brain can process information in an algorithm
* the idea has been around since 1940s, but it became a way to explain informaiton processing in the brain later on
- up until then, psychologists posited that information was processed serially; connectionist models proposed that information could be processed in parallel
connectionist model called: parallel processing model
connectionist models notes
neural networks work similar to human brain
* connectionist models are powerful in labelling things in the world in the way that we label them
* offers more functionality
* not a new idea
* improvement of idea is the power of the computer processes we have
biggest shift from other models is looking at categorization, not in a serial process, you process everything in parallel
feeding information to neural nets and telling it the output we want but the processing is done through training, we don’t specify anything
computer figures out what the crucial part of the image are – supervised training bc people are labelling images
connectionist models
- units correspond to how neurons function in the brain
- individual units may represent concepts, but this is not necessary
represents info in a more abstract way
think of these nodes as being individual units of knowledge
* knowledge is actually spread across nodes
* move away from thinking about each node and think about the collection of nodes
connectionist models – 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
more broad thinking makes it possible to represent information more mathematically and more easily
connectionist models – units
output units: receive input from hidden units
hidden units: receive input from input units
input units: activated by stimulation from environment
connectionist models – activation
- at each point in time, each unit has a degree of activation
- units feed their activation into other units
- compute activation of a unit using among of activation fed into it
the weight specifies the degree to which unit A contributes to unit B
excitatory connection: positive
inhibitory connection: negative
activation rule: specifies output activation based on input activation
the way info is spread across networks is the same way we thinking about knowledge and memory being spread across our brain
connectionist models – layers
activation determines the activation of the next layer
- concepts are represented by the pattern of activation across units
- hidden units act as abstract entities devoid of interpretation
– although this is now changing
there could be redundancies such that multiple units could represent the same concept
* this makes the model robust to “breaking” the network
conceptual interpretation are some of the things people are testing with already established and trained models
if w can understand the hidden layers, it will help us understand information being held across different levels of the brain
what are the hidden units doing?
used to have no idea but more efforts are being made to uncover the processing
they can categorize objects but we don’t know what that process looks like
connectionist models – storage
cut off connections between different layers – same as experiments with ablation
* have to break a lot of connections for the model to fail
based on parallel process, but explains how brain processes info as an algorithm
neural nets concepts are stored as distributed representations
ChatGPT
don’t know anything – just the frequent connection between words
same as other models but its size is what makes it different
it just strings words together
can’t make ideas but can put ideas together
should used it as a launching pad for what to write or how to phrase things later
how neural networks are like the brain?
- graceful degradation
- generalization
- distributed
neural networks mimic the brain quite closely
graceful degradation
- distributed representations within the network means that disrupting or breaking the system does not halt performance, only decreases it
generalizations
neural networks also have a capacity to generalize from particulars
distributed
the idea that memory and identity are distributed and redundantly stored, rather than localized and unique
cryonics: means precise reconstruction of the brain may not be necessary to restore memory and identity
concepts in the brain
neuropsychology offers insight into how concepts and categories are represented
some approaches
* sensory-functional hypothesis
* semantic category approach
* multi-factor approach
brain damage affects how some patients are able to categorize the world
different explanations trying to figure out what went wrong that produces these results
sensory-functional hypothesis
seperate semantic stores for
1. sensory or perceptual properties of objects
2. functional information related to object use
selective impairments for living and nonliving things are assumed to derive from an assymetry in the representation
representations of living things are more heavily weighted in terms of visual sensory features
representations of nonliving things are asusmed to be more heavily weighted in terms of functional features
sensory-functional hypothesis cont’d
can’t categorize different types of animals
thought of the semanted network as being interconnected
seperates knowledge: * function of the semantics and overarching differences in organization
* or problem of how the info is fed to the system
hypothesis says that the thing that really matters is how that information is processed
what matters about the sensory functional hypothesis
hypothesis says it is how the information is processed that matters
- the thing that is going wrong is that processes that largely require visual and sensory information is not being uploaded to the semantic part of the brain so its interrupted
*knowledge is accessed differently depending on which thing is weighted more
kids can categorize when they turn 6 but not before
semantic category approach
- patients with category-specific semantic deficits may have selective impairments for naming items from one category of items compared to other categories
- those patients may also have categorical impairments for answering questions about all types of object properties
trying to separate out what’s going on where is it that patients are truly having problems with
semantic category approach – looked at differently
dividing up how people percieve and if its living or not living
refuting that there’s the separation
living and non living thing is just an artifacts
its the sensory part
if it was just about the inputs it would be divided across
- because the light bars and dark bars aren’t together, it shows that the sensory functional hypothesis isn’t right
semantic category approach
Mahon & Caramazza
– little bit of sensory and semantic
- certain categories are biologically relevant for our survival, so have dedicated brain regions specialized in their processing
- Ex: faces
- they argue for distributed domain specific representations of concepts
somewhere in between:
* primary sensory and motor areas that have a physical organization in the brain that projects topographically onto a physical dimension
* distributed representation of human cognition
— abstract systems that make human thought and metacognition possible
certain neural networks are involved in responding to specific categories of stimuli
multiple-factor approach
argue that S-F approach is incorrect, as need to take into account that categories could have overlap with different features/factors
looks at how concepts are divided up within a category rather than identifying specific brain areas of networks for different concepts
propose that each category is defined by a combination of a large number of factors (In addition to the sensory/function divide)
* Ex: color, motion, action-performed could all inform non-living artifacts
middle point between the two extremes
expression of living vs non living - you could end up with this result from many different reasons
multiple-factor approach example
crowding: when different concepts within a category share many properties
ex: “animals” share “eyes”, “legs”, and “the ability to move”
started with brain damage leading to weird results
can create situations with living and non living simply by having the feature space and how much is it overlapping
more common features between items = more confusing
almost opporsite of priming
