5 - Concept Learning and Semantic Knowledge Flashcards
Semantic Approach to Knowledge
hierarchical approach
Quillian proposed that if concepts are organised into a hierarchy progressing from specific to general categories, then propositions true of all members of a subordinate category could be stored only once at a higher category
Superordinate
- propositions that describe a concept
- i.e. all plants have roots
Basic Level
- i.e. all trees have bark
Subordinate
- i.e. a Pine is a tree
To determine whether a proposition was true for a particular concept, you navigate through the hierarchy from subordinate to superordinate
Evidence of Semantic Knowledge
Collins and Quillian tested the reaction times of people asked about details of e.g. a canary:
- does it have skin
- does it have feathers
- is it yellow
Questions were answered from slowest to quickest in that order, verifying that the more basic/subordinate level information is quicker to establish, as you don’t have to compute information to go far up the hierarchy model (but this didn’t always happen)
Evaluate the Semantic Approach to knowledge
Benefits:
- Economy of storage
- Generalisation
Problems:
- Sometimes the property of generalisation within the model doesn’t work - in cases where a descriptive is not generalisable to a lower tier
i. e. birds can fly —> a penguin is a bird but cannot fly - Contrary to Quillian’s model, memory development and deterioration shows that general properties of concepts are more strongly bound to an object than more specific properties (in Quillian’s model, the more specific properties are stored closest to the object)
Potential fixes:
- store the information on flying at a lower level
- this would reduce economy of storage, since you’d have to code flying ability for each bird (most of which fly)
- contradictory information for examples below for which the above concepts are not generalisable (specifically, penguins can’t fly)
This shows that exceptions within categories are big problems for this model
Another issue with the model, is that it expects that all information at one level can be categorised into the level above at the same speed (i.e. it would take the same amount of time to define an apple as a fruit and a fig as a fruit) but this is not the case
(there is no typicality effect built in)
Prototype Approach to Knowledge and its limits
Rosch and Mervis developed this approach
based on a central description (prototype) that stands for the whole category
i.e. an apple is a fruit, if something is similar to an apple it must be a fruit
- a summary representation of a concept (defining characteristics)
- the prototypical concept might not be a real tangible thing
So the less similar a concept is to the prototype, the slower you will identify the concept as according to that prototype
- allowing for concepts that don’t quite fit with the prototype, but still fall in that category
Limits:
- it’s not always easy to think about the prototypes for a concept, especially with more complex categories
Exemplar approach to knowledge
Developed by Medin and Schaffer
- states that we make use of particular instances as examples of a category
- the more instances we have encountered of a concept, the closer (it must be) to the centre of a category (exemplar)
- this approach is able to include information about variability (unlike the prototype approach)
- using previously encountered instances in the memory, you can reason how likely a particular instance is
What are Generalisation and Economy of Storage?
Ideas that indicate why knowledge is a structured process:
Generalisation
- the tendency to respond the same way to different but similar stimuli
Economy of Storage
- storing all necessary information as economically as possible (i.e. by using an intermediate term that relates all things of the same category)
Why models of semantic knowledge are looking for generalisation and economy of storage
In terms of Generalisation:
- because propositions are stored at the highest level possible (superordinate), you can generalise that information to all of the aspects below it
In terms of economy of storage:
- by storing propositions at the highest possible level, the model stores information as sparsely as possible whilst still being relevant, giving it economy of storage
The Parallel Distributed Processing Model (PDP)
A connectionists approach
- Input Units have projections to hidden units
- These hidden units have projections onto Output Units
- Initially, all of these connections will have the same weight
- Through training (realisation of the correct output unit per input unit, by conferring) some of these connections will strengthen and some will weaken
- This change in connection weight relates to the Rescorla-Wagner equation (R-W Delta rule):
𝚫V = λ - ∑V (RW)
Change in connection:
- 𝚫W = aᵢ(correct) - aᵢ(actual)
- 𝚫W = Change in connection weight
- aᵢ(correct) = what the output should be (training signal)
- aᵢ(actual) = the output the model predicted
(error-prediction learning) - The hidden units allow a more complicated relationship between the input units and output units
What is Progressive Differentiation? (in semantic knowledge)
- Superordinate Concepts are differentiated first
- Followed by Basic concepts
- Followed by Subordinate concepts
(can be seen in children learning, and the reverse in memory deterioration)
This happens because you have more experience with the higher level concepts, the more specific the concept, the less experience you’ll have with it, thus the slower these concepts will be differentiated
What is Coherent Covariation?
- Coherent Covariation leads to Progressive Differentiation
- Items that share the same concepts (covariates) ‘cohere’ together via these concepts (i.e. birds will have wings AND hollow bones)
- Thus are separate from items that they do not cohere with
Things that cohere can be grouped together
Factors cluster together, this facilitates progressive differentiation
What do McClelland and Rogers mean by Pattern Completion and Propagation of Activation?
(within Parallel Distributed Processing {connectionist})
- Semantic information is not stored but is reconstructed in response to probes in a process called Pattern Completion
- Filling in occurs through the Propagation of Activation among units through their connections, and the outcome depends on the strengths (weights) of the connections which are shaped by experience
The structure of the Rumelhart Model
within Parallel Distributed Processing {connectionist}
- the units in one of the two input groups stand for the concepts at the bottom of the hierarchy
- the units in the other input group stand for the relations
- the output units stand for all possible completions of three-term propositions, true of the concepts
- connections are initially set to small random values so that activations produced by a particular input are weak and undifferentiated
> when the network is initialised, the patterns of activation on the representation units are weak and random, owing to the random initial connection weights, but gradually these patterns become differentiated
How the PDP model is trained
(Parallel Distributed Processing Model)
{connectionist}
Through training, weighted connections can be strengthened or weakened
- the network is trained by presenting it with experiences based on the information contained in Quillian’s hierarchy
- the hierarchy specifies that a canary can grow, move, fly and sing
> so one of the training examples specifies ‘canary’ and ‘CAN’ for the input and ‘grow’ as the target output - input units are activated
- activity propagates forward through the hidden units to the output units
- the activations resulting in the output are compared to the correct version
- the connection weights are adjusted to reduce the difference between the target and the obtained activations
What is a hidden unit?
- within the Parallel Distributed Processing Model
- a special set of internal or hidden units labelled ‘representation units’ is included between input units for the individual concepts
- hidden units represent the differentiated patterns of activation
> units in a neural network that mediate the propagation of activity between input and output layers - the activation of target values of such units are not specified by the environment, but instead arise from the application of a learning procedure that sets the connection weights into and out of the unit
According to McClelland and Rogers, why do infants develop basic level (intermediate level) descriptions first? (PDP)
- there is a clustering of objects in the world into tight-knit intermediate-level groups within superordinate categories
- there is a tendency of parents to use intermediate-level words (basic-level) more frequently than more general or more specific words when speaking to children
- a few items (such as dogs) are discussed far more frequently than more general or more specific words when speaking to children
