Week 2 – Similarity & Analogy:

Question 1

what is Similarity

Answer 1

• Similarity describes a sense of sameness.
  -We can examine values in the confusion matrix to learn about what participants were attending to.
  -some factor that can evoke similarity: Shared perceptual feature, Shared goals, conceptual and common structure

Question 2

Why study similarity?

Answer 2

It is an important part of cognition and can be apply in different area
® Learning – the transfer of learning depends on the similarity of the new situation to the original learning situation (Osgood, 1949).
® Gestalt perception –similar things tend to be perceptually grouped.
® Memory – the likelihood of remembering depends on the similarity of the retrieval context to the original encoding context (Roedigger, 1990).
-Generalization
-Categorisation – the likelihood of assigning a category label to some instance depends on its similarity to other members of the category
® Eyewitness identification: the similarity of foils to suspects in a police line-up can affect the proportion of correct identifications & false alarms

Question 3

How can we measure similarity

Answer 3

-Similarity can be measured using confusability. Items that are similar are more likely to be confused with one another when making an identification decision
-Rating scales (Likert scales)
-Respond time: how quickly can you tell.
-Force choice: which 2 is more similar in a group.
-Stimulus Arrangement: put the object together based on how similar they are on a line, the further apart, the less similarity.
-each of these methods are effective but have different pros and cons and is used for different situation.

Question 4

Experiment of similarity and rocks

Answer 4

-360 images of rocks, too many to show all at once
-Some pair have high similarity while other have low similarity
-raitings obtained from 129000+ pairs. each participant completes a random selection of pairs.

Question 5

Why measure similarity

Answer 5

-similarity measurement can be used to visualise the ‘psychological’ relationship between object
- The uncovered dimensions can be used to interpret the dimension people is seeing
-use similarity to see the representations that people hold for different stimuli
-psychological representation can be used as the basis for other theory

Question 6

3 theory of similarity

Answer 6

•Geometric models: Similarity as distance
•Tversky’s Contrast model: Similarity as commonalities and differences
•Structural Alignment: Similarity as relations

Question 7

Geometric models

Answer 7

-similarity as captured by the distance between objects. Similar things are placed close together; dissimilar things are placed far apart.
- Objects are represented in a ‘psychological space’ in which perceived similarity is represented by distance.
-The psychological space is imbued within a set of meaningful dimensions (e.g. age and adiposity) that organise the stimuli.
-the psychological representation can also be very different from the physical representation.

Question 8

Evidence for Geometric Models

Answer 8

-Similarity and Generalisation of object. the more similar an object is, the more likely we are to generalise

Question 9

Problems with geometric models.

Answer 9

o The symmetry constraint – if similarity is distance, then the similarity of A to be must equal the similarity from B to A, because the distance from A to B equals the distance from B to A.
-The triangle inequality – the distance from A to C has to be less than the sum of the
distance from A to B plus B to C.

Question 10

Geometric models must obey 3 axioms:

Answer 10

• Minimality – the distance between an item and itself should equal 0.
• Symmetry – the distance from X to Y should equal the distance from Y to X.
• The triangle inequality - Let d(a,b) equal the distance from point a to point b,
then the triangle inequality refers to the fact that d(x,y) <= d(x,z) + d(z,y)

Question 11

Feature-based models

Answer 11

-similarity is not constrained by the properties of distances. But rather shared feature. Features may be concrete (e.g. having a tail) or abstract (being derpy).

Question 12

Sub theory of feature base model

Answer 12

o Common elements model: simply counts the number of common features. The extent to which similarity increases is determined by the weight of each of those common features.
o Tversky’s contrast model: assumes that the estimation of similarity is not only affected by the number of common elements but also the number of unique elements. The unique elements of an object may exert a stronger influence on your estimation of similarity depending on where you place your focus of attention. The contrast model can predict asymmetric similarity by heightening the distinctive features of an object or by heightening the similar features of an object

Question 13

Commonalities between geometric & feature models

Answer 13

-Geometric models create hypothetical representations (dimensions) that explain the systematicities present in a set of similarity data.
-Modern models of categorization often employ this distance-based representation, but go one step further & draw upon ideas from the contrast model in that they allow the model to selectively weight different dimensions. Some models have also incorporated a dissimilarity mechanism in a manner proposed by the contrast model which allows them to capture things such as asymmetric similarity.
• Both use relatively unstructured representations – entities in objects are either sets of features or dimensions with no real relations between those attributes. However, this isn’t consistent with how we might imagine the representation of certain, more complex elements (e.g. stories, landscapes)

Question 14

Problems with feature-based models

Answer 14

• Goldstone & Yee (2005): similarity can require the conjunction of features to capture how the features are related to one another – which is not captured by either of the geometric or contrast models.
• To represent complex scenes comprised of multiple parts, a feature-type representation is going to have to get very large, very quickly

Question 15

structural aligment

Answer 15

mapping similarity between objects based on common structure. Neither geometric nor feature models take into account the fact that representations of similarity can be embedded in other representations.
o Matches in the same places (matches in place/MIPs) evokes a sense of similarity between images (e.g. having the sun in the same corner of two pictures).
o If structure is important for similarity, then MIPs should increase similarity more than matches out of place (MOPs). However, matches out of place (MOPs) should still increase perceived similarity over and above completely different features
o Goldstone (1994) revealed that matches in place (MIPs) increase perceived similarity more than matches out of place (MOPs) providing evidence for relational processing/structural alignment as being a necessity for assessing similarity between objects

Question 16

Similarity as transformation

Answer 16

-the similarity between two entities is a function of the
complexity required to “distort” or “transform” the representation of one into the representation of the other. The simpler the transformation distorting one
representation to the other, the more similar they are assumed to be.
o Goldstone found that transformations account for 94% of variance in similarity. As the number of transformations required increases, the rating of similarity decreases in a strong linear fashion

Question 17

Evidence for the distinction between similarity and categorisation

Answer 17

-Rips (1989) asked people to consider two circles or different diameter and then imagine a circle whose diameter is half-way in between such that it was equally similar to both. People were then told that the small object was a quarter and the large object was a pizza. People then overwhelmingly classified the middle object as a pizza, demonstrating that similarity (in this case based on physical diameter) was
not the sole driver of categorisation. . Other factors, such as the typical variability in size of coins and pizza were more
relevant
- Rips (1989) showed that category judgements seem to depend on biological, genetic and historical knowledge, whilst similarity judgements seem to depend more on gross visual appearance (in this experiment a bird was transformed into an
insect via toxic waste, but was still judged by participants to be a bird)

Question 18

Similarity exert a strong influence on learning

Answer 18

• Little et al., a experiment where people learn a list of word, these word are categorised in a group but there is a wrong one in each cat.
  -unless it is revel, people would still associate the wrong

Question 19

stimulus arrangement

Answer 19

the idea that similarity can be capture by distance underlies the geometric similarity

Question 20

Compromise effect

Answer 20

introducing option C (a 3rd option) increases preferences for option B.
® Theories of cognition & decision making are designed to account for data patterns that no one actually displays (data patterns that arise from simply averaging the data), and therefore those models don’t represent a true understanding of behaviour (as per Liew, Howe & Little, 2016). Furthermore, average data patterns can be robust & replicable, but are ultimately meaningless as they don’t represent any one individual

Question 21

Recommender systems

Answer 21

-are an algorithm that attempts to tailor information to each individual. They use a technique called collaborative filtering. Recommender systems rely heavily on measures of similarity in order to determine what recommendation to make next. Problems with recommender systems include that they require lots of data to generate an accurate prediction, concerns over privacy & bias (echochamber effect – amplifying prior beliefs), and the potential for a filter bubble