Categorization and concept formation

1
Categorization and concept formation
C82LEA Biology of learning and memory Lecture 15

• Charlotte Bonardi

2
Comparative psychology Concept learning Number
Time Conditional learning Just more stuff that
animals can do? NO! Associations are for all
of us clever! Dont just glue things that occur
together sensitive to correlations can track
causal relationships in man can do clever stuff
connectionist networks language, pattern
recognition so can animals do clever
stuff too?
3
Comparative psychology Concept learning Number
Time Conditional learning Asks -- How do
animals do these things? Is it the same way as
humans do them? Need to analyse tasks to ask
these questions Learn more about our abilities
too
4
(No Transcript)
5
(No Transcript)
6
(No Transcript)
7
Concept The mental representation of something.
Many concepts are based on the sharing of common
properties by items in a class.Concept
formation The induction of concepts that divide
items into classes according to their shared
properties (categorization).e.g. klee (n)
Alive small, round, furry animal that moves by
bouncing comes in various shades of red.
Frequently found in lecture theatres.
8
Note that concepts are not always defined by
specific features... sometimes do not have
NECESSARY or SUFFICIENT features to define
them... "polymorphous"e.g. What
is the defining feature of a game?
9
(No Transcript)
10
(No Transcript)
11
even simple things like tables...
12
even simple things like tables...
13
Types of concept (i) Basic level concept --
based on similarity of perceptual qualities
(e.g., bird, flower).(ii) Superordinate concept
-- groups of basic level concepts not based on
perceptual similarity (e.g. politician,
tools).(iii) Abstract concept -- does not refer
to individual entity, but to some property,
relation or state (e.g., sameness, truth).
14
Questions.. Can animals form basic level
concepts? superordinate concepts?abstract
concepts? If so, how do they do it? Do
animals form concepts in the same way as humans?

15
Basic level concept formation in animals
Bhatt, Wasserman, Reynolds Knauss, 1988 Pigeons
in a chamber with choice of four response
keys. Shown pictures of flowers, cars, people and
chairs
Birds learned to peck different keys for
exemplars of each of the four categories of
picture.
chair
people
car
flower
16
(No Transcript)
17
Basic level concept formation in animals
Bhatt, Wasserman, Reynolds Knauss, 1988 Pigeons
in a chamber with choice of four response
keys. Shown pictures of flowers, cars, people and
chairs
Birds learned to peck different keys for
exemplars of each of the four categories of
picture. Then they tested them with some new
exemplars, that they had never seen before......
18
(No Transcript)
19
Basic level concept formation in animals
Bhatt, Wasserman, Reynolds Knauss, 1988
They also were able to respond correctly to the
new exemplars, that they had never seen before.
This suggests that the birds had formed a
concept of flowers, cars, people and
chairs. However, performance was more accurate
with the training stimuli (80) than with the
novel, test stimuli (60).
20
Theories of basic level concept formation -- how
do they do it?
(i) Exemplar theory Learn about every instance
independently. Can classify novel exemplars on
the basis of similarity to learned instances.
(ii) Prototype theory Abstract a prototype that
corresponds to the central tendency of training
exemplars. Classify novel exemplars on basis of
similarity to protype.
21
Animals are clearly storing information
about the training exemplars -- which is why they
were more accurate with them than with the
novel test stimuli. This implies their
performance can be explained by exemplar theory
Theories of basic level concept formation -- how
do they do it?
22
Animals are clearly storing information
about the training exemplars -- which is why they
were more accurate with them than with the
novel test stimuli. This implies their
performance can be explained by exemplar
theory However, it is reported that humans show
the prototype effect (e.g., Homa et al., 1981)
-- they categorize the prototype more accurately
than the training stimuli, even if it has never
been seen before This is more consistent with
prototype theory ... and doesnt really fit
exemplar theory
Theories of basic level concept formation -- how
do they do it?
23
So do animals store examplars and humans a
prototype? Are humans and animals forming
basic-level concepts in different ways? Red
Rag to a Bull! so someone tries to show a
prototype effect in animals...
Theories of basic level concept formation -- how
do they do it?
24
Aydin Pearce, 1994. The prototype effect in
pigeons
A
B
C
D
E
F
25
Aydin Pearce, 1994. The prototype effect in
pigeons
A
B
C
D
E
F
The positive and negative prototypes are defined
as ABC and DEF...
ABC
DEF -
26
ABC
DEF -
The birds trained on three-element displays,
created by distorting the prototypes (swapping
one prototype element for one from the other
category)



27
ABC
DEF -
The birds trained on three-element displays,
created by distorting the prototypes (swapping
one prototype element for one from the other
category)



28
ABC
DEF -
The birds trained on three-element displays,
created by distorting the prototypes (swapping
one prototype element for one from the other
category)
--
--
--
29
ABC
DEF -
The birds trained on three-element displays,
created by distorting the prototypes (swapping
one prototype element for one from the other
category)
--
--
--
30
The animals were taught that the three positive
patterns were always paired with food, whereas
the three negative patterns were not. Birds
pecked a response key more at positive than at
negative patterns.
31
The animals were taught that the three positive
patterns were always paired with food, whereas
the three negative patterns were not. Birds
pecked a response key more at positive than at
negative patterns. Then the birds were tested
with the training patterns and the prototypes....
32
(No Transcript)
33
The birds responded more to the positive
prototype ABC than to any of the positive
patterns, and less to the negative prototype,
DEF, than to any of the negative
patterns. This is evidence of a kind of
prototype effect (though not everyone thinks
this evidence is good enough - pigeons fail to
learn more complex prototypes)
34
Narrowing the gap... humans and animals more
similar than we thought... so lets ask the
converse question do humans store
exemplars as well?
35
Whittlesea, 1987 But do humans store information
about the training items...? Prototype
FURIG 1 FEKIG FUTEG PURYG
FYRIP KURIT 2 FYKIG FUTYG PUREG FERIP
PURIT 3 FUKIP PUTIG FURYT FYREG
KERIG
36
But do humans store information about the
training items...? Whittlesea, 1987 Prototype
FURIG 1 FEKIG FUTEG
PURYG FYRIP KURIT STUDY 2 FYKIG
FUTYG PUREG FERIP PURIT 3 FUKIP PUTIG
FURYT FYREG KERIG
Pretest with all stimuli (30ms presentation
followed by a mask
then had to write down as many letters as they
could). Studied list 1 Tested with lists 1,
2 and 3.
37
Whittlesea, 1987 Prototype
FURIG 1 FEKIG FUTEG PURYG FYRIP
KURIT 2 FYKIG FUTYG PUREG FERIP
PURIT 3 FUKIP PUTIG FURYT FYREG KERIG
If they have abstracted the prototype, then they
should be equally good at categorising 1, 2 and
3, as they all differ from the prototype by two
letters...
38
So do we abstract a prototype? maybe
not... Whittlesea, 1987 Prototype
FURIG 1 FEKIG FUTEG PURYG FYRIP
KURIT 2 FYKIG FUTYG PUREG FERIP
PURIT 3 FUKIP PUTIG FURYT FYREG KERIG
But if they are remembering the exemplars, they
should be best with 1(studied) and better at 2
than at 3 list 2 more similar to list 1 than
list 3.
39
Prototype theory says subjects should be equally
good at lists 1, 2 and 3 -- all equally
similar to the prototype. Exemplar theory says
list 1 should be easiest (studied), then list
2 (differs a little from studied list) and
then list 3 (differs a lot from studied
list) And that is what they found... 1 1.07 2
0.80 3 0.51 (improvement from pretest)
40
Conclusion Both humans and animals retain
information about the training items/exemplars
(consistent with exemplar theory) So what
about prototype theory? It turns out that
exemplar theory can even explain the prototype
effect! The two theories do not make very
different predictions after all. How? Lets
go back to Pearces experiment we need to
explain why birds peck more at positive prototype
than to other members of the positive category
41
Prototypes
A B C
D E F -
Training stimuli



--
--
--
42
Positive Trained stimulus
Positive Prototype
A B C
Training stimuli



--
--
--
43
Positive Trained stimulus
Training stimuli



--
--
--
44
Positive Trained stimulus
Training stimuli



--
--
--
45
Positive Trained stimulus
five positive, four negative
Training stimuli



--
--
--
46
Positive Prototype
Positive Trained stimulus
five positive, four negative
Training stimuli



--
--
--
47
Positive Prototype
Positive Trained stimulus
five positive, four negative
Training stimuli



--
--
--
48
Positive Prototype
Positive Trained stimulus
six positive, three negative
five positive, four negative
Training stimuli



--
--
--
49
Conclusion If exemplar theory assumes that each
stimulus is composed of a set of elements, that
are more or less associated with
category membership, then it can explain the
prototype effect This explanation is actually
viewed as a new theory "feature theory"
50
Feature theory versus Exemplar theory The
difference between feature theory and exemplar
theory is subtle. They both say you store
something about the stimuli on each
trial Exemplar theory implies that each
stimulus is a configuration -- new stimuli
classified on basis of similarity to stored
configures Feature theory says that each
stimulus is composed of a set of elements, --
new stimuli classified on basis of sharing stored
elements
51
They can probably both explain the prototype
effect (but easier to show with feature theory).
There is still controversy about which of the
three theories is best
52
Some people argue that actually categories are
formed by means of associative learning- for
example, about stimulus features The features
of the category become ASSOCIATED with the
category label Features --gt Category
53
To do this they showed that these associations
are subject to
BLOCKING because this is a key
characteristic of associative learning Featur
e--gtCategory X Feature--gtCategory
X? small cr
X Feature--gtCategory X? BIG
CR Strength of association not determined by
number of pairings, but by SURPRISINGNESS OF
OUTCOME
54
An experiment by Shanks (1990), based on Gluck
and Bower (1988). Subjects given series of
trials in which sets of medical symptoms and a
diagnosis are presented. The subjects must
estimate the extent to which a particular
symptom is associated with a particular disease.
runny nose headache flu
TRAINING PHASE
55
An experiment by Shanks (1990), based on Gluck
and Bower (1988). Subjects given series of
trials in which sets of medical symptoms and a
diagnosis are presented. The subjects must
estimate the extent to which a particular
symptom is associated with a particular disease.
If I have a runny nose, what have I caught?
TEST PHASE
56
There were two diseases, one common (e.g. flu),
which occurred a lot, and one rare (e.g. pig
disease), which didnt occur often. Three
possible symptoms one target symptom (a --
e.g, a headache) and two nontarget symptoms
b, -- e.g., a runny nose, and c -- e.g.,
grunting The following is a simplification of
Shankss experiment
57
pig ac pig ac pig ac pig ac pig ac
pig ac pig c pig c pig c pig c
flu ab flu ab flu ab flu ab flu ab
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu ab flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
Test a?
(i) lots more b--gtflu pairings than c--gtpig
pairings -- 30 versus 10
58
pig ac pig ac pig ac pig ac pig ac
pig ac pig c pig c pig c pig c
flu ab flu ab flu ab flu ab flu ab
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu ab flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
Test a?
(ii) a always presented with b on flu trials,
and with c on pig trials
59
pig ac pig ac pig ac pig ac pig ac
pig ac pig c pig c pig c pig c
flu ab flu ab flu ab flu ab flu ab
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu ab flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
Test a?
(iii) same number of ab--gtflu and ac--gtpig
pairings -- 6 and 6
60
pig ac pig ac pig ac pig ac pig ac
pig ac pig c pig c pig c pig c
flu ab flu ab flu ab flu ab flu ab
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu ab flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
Test a?
If number of pairings are critical subjects
would, given symptom a be equally likely to
diagnose flu as pig-disease - same number
of pairings....
61
pig ac pig ac pig ac pig ac pig ac
pig ac pig c pig c pig c pig c
flu ab flu ab flu ab flu ab flu ab
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu ab flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
Test a?
Associative learning makes a different
prediction
62
Associative analysis
of task Subjects are given symptom a, and asked
to say which disease it predicts, flu or pig
disease. Thus the critical factor is the
strength of the a---gtflu and a--gtpig
associations -- which is stronger? Symptom a is
paired with flu the same number of times as it
is paired with pig disease, so all other things
being equal, a should be associated equally
strongly with both diseases.
63
pig ac pig ac pig ac pig ac pig ac
pig ac pig c pig c pig c pig c
flu ab flu ab flu ab flu ab flu ab
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu ab flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
Test a?
But all other things are not equal a is not
paired alone.... ab--gtflu and
ac--gtpig
64
pig ac pig ac pig ac pig ac pig ac
pig ac pig c pig c pig c pig c
flu ab flu ab flu ab flu ab flu ab
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu ab flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
Test a?
. .. and b predicts flu better than c predicts
pig more pairings of b?flu
65
pig ac pig ac pig ac pig ac pig ac
pig ac pig c pig c pig c pig c
flu ab flu ab flu ab flu ab flu ab
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu ab flu b flu b flu b flu b
flu b flu b flu b flu b flu b
flu b flu b flu b flu b flu b
Test a?
b
c
pig
flu
66
The Rescorla-Wagner model predicts that an
association will only form between two events if
the second event is surprising. That is why you
get blocking. Thus in normal conditioning

a
flu
get good learning about a because flu is
surprising. But in blocking
b
a
flu
ab
flu I know...
flu
get poor learning about a because flu is NOT
surprising -- it is predicted by b. This is
blocking.
67
In the diagnosis task b--gtflu association is
strong. Because b is a good predictor of flu,
the a--gtflu association will be blocked
b
a
flu
ab
flu I know...
flu
68
But c--gtpig association weak c--gtpig As c is a
poor predictor of pig, less blocking occurs
c
pig
ac
pig ?!! really?!
a
pig
69
Exemplar theory predicts that, given symptom a,
subjects will be equally likely to predict flu
as pig disease. Associative theory predicts
that, given symptom a, subjects will be far more
likely to choose the rare disease, pig disease.
70
Exemplar theory predicts that, given symptom a,
subjects will be equally likely to predict flu
as pig disease. Associative theory predicts
that, given symptom a, subjects will be far more
likely to choose the rare disease, pig
disease. Associative theory wins! Proportion
common disease (flu) diagnoses .37 Proportion
rare disease (pig) diagnoses
.63 Suggests that associative learning is the
best explanation of performance on this
categorization task in human subjects.
71
Superordinate level concept formation
72
(No Transcript)
73
Superordinate level concept formation in animals
Superordinate categories have members that are
not necessarilyphysically similar to each other,
but share a common associate.Can animals can
form categories of this kind?
74
Wasserman, De Volder Coppage, 1992 Pigeons
trained with slides of people, chairs, cars and
flowers
The birds reinforced for making R1 (pecking one
specific key) to either people or chairs, and for
making R2 (pecking a different key) to either
cars and flowers. people and chairs in one
category, and cars and flowers in another.Then
trained to make R3 to people, and R4 to
cars.Tested with chairs and flowers, and given a
choice of R3 and R4. R3 to chairs and R4 to
flowers counted as correct responses.
75
(No Transcript)
76
Superordinate level concept formation in animals
Animals seem to have formed a superordinate
category treating people and chairs as
equivalent because both paired with the same
response in the first phase of trainingThis is
a more complex type of categorization because the
categorymembers are not physically similar.Is
this the same as how people do it?Some people
(e.g. Pearce, 1997) argue that this is not true
categorization like that seen in humans, but just
simple associative learning -- and that what we
do is somehow more complicated. But need to
specify exactly how what we do is more
complicated, so we can test...
77
Abstract concept formation
in animalsRelatively little work has been done
on abstract concepts in animals. The one that
has been studied most is same/different.Usually
studied using a Match-to-sample technique (MTS).
Birds shown a sample key e.g. red then they
were given a choice of two comparison keys, red
and green. Must peck the same colour as the
sample -- i.e. red. On other trials the bird
gets a green sample then task is to peck the
green comparison.
78
(No Transcript)
79
correct
80
correct
81
correct
correct
82
Birds could master this, but were typically poor
at transferring toskill to different colours
(e.g., blue and yellow). This suggests they had
not really learned the concept of same.
correct
correct
83
However, more complex training techniques seemed
to produce better results Wasserman, Hugart
Kirkpatrick-Steger, 1995. Pigeons were shown
complex stimulus displays, and given a choice of
a red and a green key.
84
(No Transcript)
85
Abstract concept formation in animals
They were reinforced for pecking the red key on
same trials, and the green key on different
trials. They were trained in this way for
arrays involving one set of specific icons (the
top half of the figure)Finally they were tested
with a further set of arrays involving a second
set of specific icons (the bottom half of the
figure).
86
(No Transcript)
87
(No Transcript)
88
ReferencesGeneral Bouton, M.E. (2007)
Learning and Behavior. Sinauer Associates.
Lea, S.E.G. ((1984). In what sense do pigeons
learn concepts? In Animal Cognition (Roitblat,
H.L., Bever, T.G., Terrace. H.S. (Eds.)
(pp.263-276). Lawrence Erlbaum
Associates.Pearce, J.M. (1997). Animal
Learning and Cognition. Lawrence Erlbaum
Associates.Wynne, C.D.L. (2001) Animal
Cognition. Palgrave MacmillanShanks, D.R.
(1995) The Psychology of Associative Learning.
Cambridge University Press. http//www.pigeon.ps
y.tufts.edu/avc/huber/
89
Aydin, A., Pearce, J.M. (1994). Prototype
effects in categorization by pigeons. Journal
of Experimental Psychology Animal Behaviour
Processes, 20, 264-277. Bhatt, R.S.,
Wasserman,E.A., Reynolds, W.F., Knauss, K.S.
(1988) Conceptual behaviour in pigeons.
Journal of Experimental Psychology Animal
Behaviour Processes, 14, 219-234. Gluck, M.A.,
Bower, G.H. (1988). From conditioning to
category learning an adaptive network model.
Journal of Experimental Psychology General, 117,
224-247. Homa, D., Sterling, S., Trepel, L.
(1981). Limitations of exemplar-based
generalization and the abstraction of
categorical information. Journal of Experimental
Psychology Human Learning and Memory, 7,
418-439. Shanks, DR. (1990). Connectionism and
the learning of probabilistic concepts.
Quarterly Journal of Experimental Psychology,
42A, 209-237. Wasserman, E.A., Hugart, J.A.,
Kirkpatrick-Steger, K. (1995). Pigeons show
samedifferent conceptualization after training
with complex visual stimuli. Journal of
Experimental Psychology Animal Behaviour
Processes, 21, 248-252. Wasserman, E.A., De
Volder, C.L., Coppage, D.J. (1992).
Non-similarity-based conceptualisation via
secondary or mediated generalization.
Psychological Science, 3, 374-379. Whittlesea,
B.W.A. (1987). Preservation of specific
experiences in the representation of general
knowledge. Journal of Experimental Psychology
Learning, Memory Cognition, 13, 3-17.
90
Lectures and handouts on my webpage webquiz for
multiple choice practice http//www.sinauer.com/
bouton/ cmb_at_psychology.nottingham.ac.uk