You are on page 1of 5

Psychology of Thinking Concept Chart

Professor Matt Serra

Duke University - Fall 2006
Vanessa Villamia Sochat

Form of Representation Assumptions General Criticisms Experimental How does it deal with…
Criticisms Typicality Effects Unclear Nesting
Cases (disjunctive
1) representation of a concept is a summary 2) The features that represent a concept are 1) Functional features (but nothing 1) Findings reflect a categorization Shouldn’t occur under Should have
description of an entire class rather than a set of singly necessary and jointly sufficient to says can’t have them) process… we infer thinking process classical view 100% certainty
descriptions of various subsets or exemplars of define the concept 2) Cannot account for disjunctive from behavior.
Classical View that class. concepts (but do they exist?) 2) We ask people to list features… COMLPEXITY
3) Two instances of the same concept 3) Disagreement among individuals we don’t know if using non- MODEL
result of an abstraction process cannot be totally disjunctive under classical (but they don’t know features) necessary or defining – get rid of
need not correspond to a possible, specific view 4) Specification of defining features need to list features with Added intrinsic/extrinsic
instance (just because we haven’t figured them MULTIDIMENSIONAL features to deal with
applies to all possible test instances (applied to 4) Nesting of features out doesn’t mean they don’t exist! SCALING gradient of membership
anything I think is a pen) 5) DEATH NILL: Nested concepts!
Subset should be more similar to
immediate superordinate over any
distant set. This doesn't always
happen. (but again, asking ppl to
produce, may be mediated by
5) Should

Has classical features, nesting, etc We compare 1) typicality is an inverse of complexity. 1) Has limited capacity: we can only 1) Ctegorization is based on 1) a probe concept
Complexity Model features of target and probe and say no as soon as More typical = less features. More atypical do so much at one time defining features which are should be categorized
we find a mismatch = more features 2) Takes no longer to respond to relatively inaccessible faster when the target
2) We name typical instances because probe atypical than typical probe when probe 2) We can’t tell if people listing concept is a distant
Two steps: acts as memory probe and the most typical isn’t a member of the concept. necessary or non necessary features superordinate than
1) Access thing comes to mind. 3) People generate more features for Maybe they are categorizing using when it is an
2) Comparison typical than atypical concepts non necessary features mediated by immediate one
4) Family resemblance – found graded defining features? The only way to because there are
Target = category membership in cateogories test this is to use fewer features to
Probe = instance MULTIDIMENSIONAL compare
SCALING – gets ride of need to list
People rank similarity of pairs, 1) Closer things clump together based on 1) People judging similarity on
Multidimensional similarity perceptual (non-necessary features) –
2) We can also do this with features to see slaps classical view in the face!
Scaling which features cluster around which
1) People can check interconnect links 1) Paths allow for shortcuts. This An atypical member has This view was created
Salvaging Classical rather than checking features of concept model is lacking in constraints a longer path than a to get around nesting
when doing categorization facts. 2) Says nothing about non-necessary typical member, so takes issues. Says nothing
View Information comes from links. Links of features, and it’s obvious that we use more time. about disjunctive
paths will vary. Longer paths = longer them concepts
processing time.
Access Links Between
Translation of Features Rule plus exception approach: Rule is classical 1) Allows for translation from one feature 1) Has no way dealing with non Typical things require
view. Exceptions are translated (the atypical to another (flying implies animate) so drops necessary features fewer translations, so it
ones) nesting all together 2) If we add enough translation, there takes less tome.
isn’t need for defining features at all,
so we don’t even need classical view
Greater Accessibility of ID Identification features (perceptual) ARE Shifting theoretical work away from This view has No need for nesting
important, so let’s focus on them. defining featurers (core) so why bother difficulty with because we are
Features with them at all? This view salvages unclear cases – dealing with non
the classical view by throwing it away. if given enough necessary features
time should be Disjunctive Concepts:
able to figure caused by perceptual
something out stuff so it doesn’t
Has summary description, representation of a 1) Features that go into this summary 1) Just listing features to describe Typical members match unclear cases Disjunctive concepts:
concept is assumed to be results of an abstraction representation are salient (tangible, something doesn't give us enough more features than non can occur if we Since the same
process which is not necessarily realizable as an perceptual, noticeable) ones. information. Doesn't specify kind of typical, so reach sum are right at the weighted sum can be
Prototype Theory instance. But we don’t need necessary features 2) They have a substantial probability in knowledge we have about concepts. faster threshold for reached by
occurring in instances of the concept. People know about relationships membership or combinations of
3) Featural approach deals with modal between features and the variability slightly below various features, it
features (non-necessary features represented that is permissible within that concept. it, or if we reach follows that various
Featural Approach by weights, which are reflective of salience Needs a way to incorporate that threshold for feature sets can be
and probability) information! one or more used to determine
4) Categorization is determined by a different category membership.
weighted feature sum 2) What features could be plausible? concepts at the
5) no specification of defining features, we What's allowed and what isn't? same time. Nesting Problem:
are using modal features, so deals with use Someone’s feature may not be weighted sum for
of non-necessary features well someone else’s. We may have pseudo chicken
features. Features should have some matches/reaches
degree of generality. There are no animal sum more
constraints on what can be listed as quickly than bird

3) Has little if anything to say about

context. (The prime might be concept,
so spreading activation hints at
context) There's not enough
information given by discrete features.

Spreading Activation Represented as summary descriptions that 1. Presented with something, representation CRITICISM FEATURAL AND Typicalty effects: when Disjunctive concepts:
contain many non-necessary features. Each activated DIMENSIONAL APPROACHES dealing with a weighted Weighted sum could
feature is weighted by its importance in 2. Activation spreads on paths to anything sum, typical ones reach be reached by any
conferring concept membership. This is called connected to 1) Neither approaches seem to it faster than atypical variation of
criteriality. You only have to store features once, 3. Intersecting links are brought forward for represent all of knowledge contained in combinations.
and members of concepts are connected to these further evaluation. Features aren't. a concept, particularly relations
features. 4. If links share same label, information between concepts
accumulated. We accumulate matches and
mismatches. We either accumulate enough 2) Very few true constraints on these
+ evidence or enough – evidence to surpass models, who si to say what are correct
some threshold features/dimensions

Non-necessary features are built into this 3) Neither deals with context effects at
model! Defining might be built in, but it all, spreading activation hints at it, but
doesn’t matter none of them talk about it explicitly
4) Metric space may not hold true in
the brain
Feature Comparison Model There is a summary representation using 1) The person ignores all weights. They
weighted non necessary features but this does not simply determine number of feature
necessarily rely on feature sums for processing matches between test item and target
concept. Could be accomplished by
simultaneously comparing each feature of
target with each feature test item. If the
number of matches exceeds some high
criterion, the item mostly likely belongs to
the concept, and we say YES. When we are
between high and low criteria we start stage
2 processing
2) person selects only those features of the
test item and the target concept with high
weights and determines whether each
highly weighted feature of the target
matches such a feature of the test item. If
all such features match, the item is a
member of the concept, otherwise it is not
representation of a concept as a summary 1) any dimension used to represent a
Dimensional description that applies to all instances concept must be a salient one with at least
some of its values having a high probability
Approach concept representation depicts the average or of occurring in instances of the concept.
mean dimension values for the entire class 2) the value of a dimension represented in a
concept is the subjective (an interpretation)
With means we can have something that is a average of the values of the concept's
member of the concept that looks like nothing subsets or instances on that dimension
that we see every day 3) Non necessary dimensions can be
included like non-necessary features
Relationship between any two concepts or 4) Weights are also used, and each weight
instances of concepts can be judged based solely indicates the importance of variations in the
on distance estimates associated dimension of concept
5) Concepts with same relevant dimensions
can be represented as points in a
multidimensional metric space. Must have
minimality, symmetry, and triangular
Simple Distance Model relation between any pair of concepts or its 1) anything that fall in the threshold The closer item x is to Unclear cases disjunctive concepts.
instances is given by the distance between points difference is in the category concept Y the faster and are on the imit (everything has
and a multidimensional representation. 2) there is no decomposition of concepts more accurately X will (border) of the relevant dimensions,
into features. be categorized as a distance doesnt matter if is
3) The closer item x is to concept Y the member of Y. shorter threshold large or small) dont
faster and more accurately X will be distance = more typical. need same
categorized as a member of Y. dimensional values as
long as we're in a
We don’t need defining features because we relevant dimension
only care about distances.
Comparative Distance Model Simple distance model doesn't consider influence Assumption 1: An entity, x, is categorized
of contrasting concepts. We need a model that as an instance or a subset of concept Y iff
compares distance with test item (target concept) the metric distance between x and Y is less
and the distance between the test item and any than that between X and any concept that
item that might contrast with the target. contrasts with Y.

Assumption 2: The greater the difference in

the distance between X and Y on the one
hand, and x in any contrast of Y on the
other hand, the faster and more accurately x
will be categorized as a member of Y.

Holistic Approach

Concepts are represented by their exemplars, at 1) representation of a concept consists of If concepts are represented by their
least in part. Rather than by an abstract separate descriptions of some of its exemplars, is there room for
summary. exemplars either instances or subsets. abstraction at all? Often term
Exemplar View 2) Exemplars themselves can be represented exemplar is used ambiguously
Can be a specific instance or a subset of a in different ways. it partly depends on
category whether the exemplars are subsets or
specific instances.
3) If the exemplar is an instance, it must be
represented by the property description.

Proximity Model each concept is represented by all of its instances, The concept representation is lacking in Model leaves no room for abstraction
all of its instances that I have ever encountered. abstraction, what comes to mind is a Drawback: sheer amount of suff that I
Violates all three of summary representation specific exemplar that exists in the world. have to store! having to index through
assumptions Every exemplar in the representation is it all the time would slow us down, this
realizable as an instance doesn't seem logical, saying that we
the information that you retrieve when are cognitively lazy
making a decision about a particular
concept membership varies with the test
item presented
Best Examples Model way of constraining what comes to mind Any entity is categorized as an instance or a Typicality Effects: Unclear Cases: Disjunctive
Representation of a concept is restricted to subset of a category iff that entity retrieves Reach criteria faster for Can be unclear Concepts: each
exemplars that are typical of the concept: focal a criterial number of the concepts typical than atypical. case because concept representation
instances exemplars before retrieving a criterial Typical instances have fails to reach is explicity
number of exemplars from any contrasting more property overlap. criteria number disjunctive. An item
Exemplars represented are the ones that share concept. exemplars for that belongs to a
some criterial number of properties with other The probability that this entity retrieves any any concept or concept if it matches
exemplars in the concept. they have some specific exemplar is a direct function of the reaches criteria exemplar x, y, it
critical family resemblance score. similarity of that entity and the exemplar it number for two doesn't have to match
retrieves concepts at them all. Only has to
Categorization is accurate to the extent the same time match one. Doesn't
test probe is similar to the concept. have to match
This is an indictive process: we are everything. There
inducing (bringing things in and building) isnt any defining
rather than deducing. anything

non-necessary features: they are ALL

non-necessary features because this is an
exemplar approach
Context Model Exemplar based model
Abstraction based on context
Differences from the exemplar model
1) deals with learning of exemplar model
2) deals with computation of similarity in the
categorization process

Agree on categorization and conceptual


All Three
Functional differences between metric, dimensional approaches and featural approaches:

1) Representation of properties is different: continuous dimensions vs discrete features

2) Basis of abstracted representation: one uses dimensional value, other uses modal feature

3) Processing strategy is different: one is comparison of distance computation, other is sum of weights

SIMILARITIES: featural vs dimensional

1) both still probabilistic approaches - everything doesn't have to be there

2) explain use of non-necessary features in the same way and they deal with the difficulty in specifying defining properties in the same way by using
representations that require only non-necessary properties

3) Both approaches allow for degrees of disjunctiveness by permitting different combinations of properties to yield the same threshold quantity.
weighted feature sum in featural approach and distances in dimensional approach

4) Dealing with uncare cases - construe unclear cases very similarly. These are items that do not quite reach the threshold quantity or are equally
close to thresholds of one or more concepts. could be weighted feature sum or distances

5) Explain many simple typicality affects by assuming that the typicality of a member reflects how similar its properties are or how close it is to its
parent concept.

6) Data on nested triples: Assume that properties of most concepts are more similar to those of the immediate than the distant superordinate (what
classical view says) Dimensional and featural can account for exceptions because of idea of weighted sum. Chicken can reach necessary threshold
for animal before bird.

7) Both can explain the use of contrast concepts by considering the relation between a test item and the target concept and the test item and its
contrast concept. Done implicitly in featural approach:
Given these, featural and dimensional can explain a lot of the same stuff.