Semantic memory

Question 1

What is semantic memory?

Answer 1

General knowledge (no specific event).

Question 2

Anatomy of semantic dementia

Answer 2

Anterior parts of the temporal lobe.

Question 3

2 reasons for categorisation of objects and events (unit of semantic categories)

Answer 3

1. Economy (encoding of general structure).
2. Generalisation (apply old knowledge to new situations).

Question 4

Economy in categorisation

Answer 4

Pick up general pattern from individual instances. Common info is stored once, making it more efficient. Then use specific info (outstanding details).

Question 5

Generalisation in categorisation

Answer 5

Applying prior knowledge in new contexts and predict properties about a new object.

Question 6

Why are concepts so important?

Answer 6

They are like ‘ the glue that holds our mental world together’. Ties past experiences to present interactions with the world.

Question 7

What type of hierarchical structure does semantic knowledge have?

Answer 7

Taxonomic (superordinate/basic level/subordinate).

Question 8

Collins and Quillian (1969) ‘a canary…?’ true or false

Answer 8

Evidence of hierarchical storage. RTs are faster for descriptions that are specific to the concept (eg. ‘can sing’) and increase as the descriptions move further up the hierarchy. Different properties are stored at different hierarchical levels so ‘can sing’ immediately activates concept of canary while ‘has feathers’ searching until the ‘bird’ node is reached.

Question 9

Mandler and McDonough (1993) semantic development in infants habituation

Answer 9

Children learn general categories first, before making finer distinctions.

Infants habituate to toys that are different animals. 9 month-olds showed no dishabituation to the rabbit but there was a significant increase in playing time with the bus. There was no significant difference in playing time for the 11 month=olds. 9 month olds did not distinguish between different types of animals. (Same with 4x dog habituation and then probed with a dog or a fish. Significant difference only for 0;11).

Question 10

Keil (1979) predicability trees

Answer 10

Investigating which predicates could go with which concepts according to 5, 7 and 11 year-olds.

As age increases, there is a finer differentiation as age increases (a rabbit can no longer be sorry). A progressive differentiation about the features of concepts (not simply concepts themselves).

Question 11

Development of loss of concepts in semantic dementia

Answer 11

Loss of specific concepts first. More general concepts are kept.

Question 12

Acquisition and loss of knowledge progression

Answer 12

Development: acquire superordinate-level concepts first.
Semantic dementia: lose superordinate-level concepts last.

Question 13

Is there a single hierarchy of knowledge?

Answer 13

No, concepts can be flexible (dogs are both canines and pets). Concepts are context-dependent (a disease that affects dogs is more likely to affect wolves; a shop that sells dogs is more likely to sell cats).

Question 14

Rosche (1973) on within-category differences

Answer 14

Concepts cluster around prototypes. An apple is a fruitier fruit than an olive.

Question 15

What are the 4 effects of prototypes?

Answer 15

1. Faster categorisation and verification.
2. More frequently listed in a category.
3. Learnt first and spared in SD.
4. Recognised well even if never seen.

Question 16

Lupyan (2013) on how ‘all-or-none’ concepts are structured

Answer 16

Fuzzy concepts. Even for triangles, which has a strict rule on group membership, some triangles are more triangular than others.

Question 17

Rosch et al. (1976) on how basic level is different

Answer 17

Basic level has an advantage in naming (picture verification, language learning, SD). Psychologically privileged as they strike a good balance of being inclusive but informative.

Question 18

Empirical challenges in semantic knowledge

Answer 18

1. Complex knowledge structure - why are some levels/concepts more important? Hierarchical structure, typicality, basic level.
2. Developmental challenges - how is structure learnt and how is flexibility preserved?
3. Conceptual challenges - how is structure coded neurally?

Question 19

What is a connectionist model?

Answer 19

A network of simple interacting units with modifiable connections (based on experience). Input units -> hidden units -> output units.

Question 20

Function of units

Answer 20

Integration and activation.

Question 21

What is learning in a connectionist network?

Answer 21

The changes in the strengths of connections.

Question 22

5 principles of connectionist models.

Answer 22

1. Neurons integrate information (input from many other neurons).
2. Neural activity reflects level of input (more input = more activity).
3. Layered brain structure.
4. Influence via connections.
5. Learning alters connection strengths.

Question 23

What does the processing unit do? 3 steps.

Answer 23

1. Integrate input from the previous layer (input from multiple other units).
2. Transform the net input to an activity level (ai).
3. Transmits activity level to units in next layer.

Question 24

How is input in unit(i) calculated?

Answer 24

Input(i) = activity level of previous unit(j) * the strength of connection between the units(ij).

Netinput(i) = the sum of inputs.

Question 25

How is unit(i) output calculated?

Answer 25

The function is sigmoidal (x = netinput(i); y=activity level(i).

Question 26

Equation for unit output

Answer 26

a(i) = 1/(1+e^-netinput(i)). When net input is 0, activity is 0.5. Activity is bound between 0 and 1.

Question 27

How are the weights of the connections learnt?

Answer 27

Error-driven learning. The required output is specified.

Question 28

What sort of learning rule is used in connectionist models?

Answer 28

Backpropagation.

Question 29

What is the Delta rule?

Answer 29

The equation for the change in connection strength. Blame is a(j), the activity level from the previous unit. Change in weight = (correct activity - actual activity) * blame * learning rate. = error term * blame * learning rate.

Question 30

What sort of learning process happens in PDP models?

Answer 30

Gradual and iterated learning process.

Question 31

What is the backpropagation of error?

Answer 31

Activate 2 inputs -> correct output? If no -> change weights.

Question 32

What sort of input did Rumelhart (1990) use?

Answer 32

Unstructured input. 'Epochs' of input-output learning (item + relation -> features).

Question 33

What were the internal representations of PDP in Rumelhart (1990)?

Answer 33

Structure was learnt from unstructured input (emergent structure). The pattern of activity between similar items was similar. The network learnt by itself how to represent items based on overlapping structures of input.

Question 34

How are concepts represented in PDP models?

Answer 34

Hidden units show that the network learnt different concepts as distributed patterns of activity (no single unit represented a single input/output).

Question 35

Rogers and McClelland (2004) explaining the progressive loss in SD

Answer 35

Noise in the hidden layer. By adding noise in the hidden layer, a specific object is pushed out of its specific spot but remains in the broader area (so broader features are retained while specific features are lost).

Question 36

Rogers and McClelland (2004) explaining typicality effects

Answer 36

The network represents items based on similarity. Within a concept, there is a graded pattern of similarity (some items are really close and some are further away). Distance affects categorisation.

Question 37

Rogers and McClelland (2004) on why the basic level seems special

Answer 37

There is nothing special about the basic level. Instead, it is due to the effects of item frequency (parents use the basic level more often). Frequency affects connection strength. Learnt quickly and resistant to noise/damage. (Model can't address why the basic level is more frequent).

Question 38

Strengths of PDP

Answer 38

1. Simple. 2. Economy and generalisation. 3. Graceful degradation (due to distributed processing). 4. Represents both rules and exceptions.

Question 39

3 ideas of theory of semantic knowledge

Answer 39

1. Coherent covariation. Learn to associate objects with properties that tend to occur together. Statistical regularity help learns categories. 2. Concept similarity in a multi-dimensional space. Similarity = how alike activation patterns are. 3. Knowledge structure via learning. Iterative process of error-driven statistical learning.

Question 40

PDP efficiency

Answer 40

Capacity to store more information if needed because n units = 2^n items stored.

Question 41

PDP generalisation

Answer 41

Once network learns that 'sparrow is a bird', the network is able to generalise knowledge based on prior knowledge of other bird types.

Question 42

PDP fault tolerance

Answer 42

Graceful degradation. Distribution = no single unit is crucial. Complex representation = noise isn't catastrophic.

Question 43

PDP rules and exceptions

Answer 43

Can be simultaneously represented. Representations are 'graded' (contrast with dual-route theories).

Question 44

3 criticisms of semantic PDP

Answer 44

1. Model has built-in features and is told the right answer. 2. We know more than similarity - we also understand causality. 3. Does not explain how knowledge is actually used (combine concepts/draw inferences).

Question 45

3 general criticisms of PDP

Answer 45

1. Learning methods aren't plausible. Backpropagation isn't biologically realistic. 2. Fragile new learning overrides old connections. 3. They are not really cognitive models (just simple statistics/too underconstrained and no rules).