1. EP2.
Social learning
Elena Pasquinelli
Educa4on, cogni4on, cerveau
Cogmaster 2010‐2011
2. Transmission of generic knowledge
• Induc4on problem: Humans are capable of
transmiFng/extrac4ng general knowledge from
par4cular instances
– When such instances are repe44ve and frequent,
sta$s$cal mechanisms* are invoked.
– When this is not the case (single instance) we need a
further mechanisms for explaining induc4on.
• Such a mechanisms is hypothesized to rely on
human‐human communica4on
– Verbal and not verbal (demonstra4on)
4. Learning = the modifica4on of behavior in light
of experience
• sta4s4cal learning, • Under this defini4on, learning is a
common func4on to different animal
• learning by imita4on, species
• explana4on‐based or
causal learning
• and learning by analogy.
• Using these simple
learning mechanisms, the
brain appears to build up
complex representa4ons
about how the world
is.” (Goswami, 2008, p.
52)
5. Early learning mechanisms
1. sta4s4cal learning
• “Babies appear to be able to make
connec4ons between events that
are reliably associated, even while
in the womb.
• Once outside the womb, they
appear to be able to track sta$s$cal
dependencies in the world, such as
condi4onal probabili4es between
visual events or between sounds.
This turns out to be a very powerful
learning mechanism.”
6. Sta4s4cal learning and language
• Language acquisi4on has provoked a
debate on nature (Chomsky) vs nurture
(Skinner)
• Cri4cal periods in language learning differ
in the three aspects of language:
phone4cs (before 12 months), syntax
(18‐36), lexicon (forever)
• Why are children be^er than adults?
• Kuhl, 2004: neural commitment
– Once perceptual systems are commi^ed
they filter new informa4on
– Commitment is done between 6 and 12
months (for phone4cs): before, children
dis4nguish all the phone4c units of all
languages
• How can children succeed in a difficult
task as iden$fying and grouping the
more or less 40 phonemes that compose
their language? In the middle of a great
variability of speech?
7. Sta4s4cal learning and language
• Sta4s4cal learning
(Saffran, et al, 1996)
applies to the capacity
to iden4fy phonemes
and to the capacity of
segmen4ng words
– Japanese and English
infants are both
exposed to both /r/
and /l/ sounds, but in
Japanese the sound /
r/ is much more
frequent
– Babies spot the
transi4onal
probabili4es between
syllables
8. Language: sta4s4cal learning is not
enough
• Sta4s4cal learning can have
strong and durable effects on
phone4cs at 9 months of age, and
with short‐4me exposure to
sta4s4cal regulari4es
– 9 months old children can learn to
dis4nguish Mandarin phonemes
from exposure to play and
interac4on with a Mandarin
speaking tutor
• But is sta4s4cal learning enough?
– 9 months old children cannot learn
to dis4nguish Mandarin phonemes
from a Mandarin speaking TV‐
canned /audiotaped tutor
• Social interac4on is required
9. Social interac4on
• Social interac4on can have
an effect on learning
through:
– Enhancement of a^en4on
– Addi4onal informa4on
(gaze to object)
– Ac4va4on of mirror
systems, and other
mechanisms for
percep4on‐ac4on linking in
the brain
10. Implicit learning
• Implicit learning theories are based
on the capacity of extrac4ng
regulari4es, e.g. on grammar:
– Reber, 1967, 1989: implicit learning
allows the acquisi4on of complex,
abstract knowledge without awareness
and effort (extrac4on of abstract rules)
– Pacton & Perruchet, 2006: acquisi4on
of the ap4tude to correctly answering
to certain situa4ons, without the
inten4on of learning (no extrac4on of
abstract rules; the learning of rules
requires explicit learning)
• the crucial variable is the exposi4on
to regulari4es in the environment
12. Implicit & explicit learning
• Perruchet & Pacton, 2006: Explicit learning
completes implicit learning with rules
• Perruchet & Pacton, 2006: In any case,
explicit learning raises performances in
comparison with implicit learning (school
instruc4on demands more than above
chance performances)
• Reber, 1989: introduc4on of explicit
instruc4on is especially useful when
informa4on is provided before (rather than
during or aker the implicit learning phase),
maybe because it helps direc4ng a^en4on
on meaningful aspects
• Bransford, Brown, & Cocking, 2000: Judd &
Scholckow 1908’s experiment confirms
that explicit instruc4on (before training)
enhances performances for new situa4ons
13. Implicit learning of errors
• If implicit learning can
happen by repeated
exposi4on (with
a^en4on), then the
repeated exposi4on to
errors favors the learning
of errors
• Mul4ple choice tests
enhance learning of good,
and bad, answers (Marsh,
et al., 2007, p. 195)
14. Sta4s4cal learning & Extrac4on of
causal structures
• “… specific perceptual features
of two objects in a “launching”
event (where object A impacts
object B, causing it to begin to
move) may vary, but spa4o‐ h^p://cogweb.ucla.edu/Discourse/
temporal dynamics (and Narra4ve/micho^e‐demo.swf
therefore causal structure, i.e.,
the fact that A causes B to
move) will vary less. The
perceptual “illusion” of
causality during launching and
other visual events noted by
Micho^e (1963) is one example
of how perceptual covaria4on
can yield causal (Goswami,
2008b, p. 9)
15. Early learning mechanisms
learning by explana4on & analogy
• “In the field of machine
learning, explana4on‐based
learning depends on
construc4ng causal
explana4ons for phenomena on
the basis of specific training
examples, using prior domain
knowledge.
• If infants were merely learning
condi4on‐outcome rela4ons, as
in associa4ve learning, then
they would be unable to make
predic4ons about novel
events.” (Goswami, 2008, p. 66)
16. Learning by analogy
• “In learning by analogy, “we face a situa4on, we
recall a similar situa4on, we match them up, we
reason, and we learn” (Winston, 1980). We may
decide whether a dog has a heart by thinking about
whether people have hearts (young children use
“personifica4on analogies” to learn about biological
kinds, see Inagaki & Hatano, 1988), or we may solve
a mathema4cal problem about the interac4on of
forces by using an analogy to a tug‐of‐war (young
children use familiar physical systems to reason
about unfamiliar ones, see Pauen, 1996).
• Reasoning by analogy has usually been measured in
children aged 3 years or older (see Goswami, 1992,
2001, for reviews), but can also be demonstrated in
infancy. However, so far, analogy has not been found
in the animal kingdom, sugges4ng that it is
especially important for human
learning.” (Goswami, 2008b, p.13‐14)
17. Early learning mechanisms
Imita4on
• “Learning by imita4on can be
defined as B learns from A some
part of the form of a behavior…
• One example is learning the use
of a novel tool by imita4ng the
ac4ons of another user with that
tool. (Goswami, 2008, p. 62‐63)
Learning by imita4on is present in the
human baby by the age of at least 9
months (Meltzoff, 1988) Meltzoff, 1988
18. Learning by imita4on & TV
• 14 months’ babies can learn the same ac4ons
from real experimenters and from experimenters
canned in a TV video (on live)
• But they learn less than from live ac4on (video
deficit effect) (Zack, et al. 2009, p. 14)
– Maybe because the processing of 2D s4muli is
poorer than the processing of 3D s4muli
– Or because 2D s4muli are poorly understood and
their rela4on to 3D real objects is not granted
– Or because of poor representa4onal flexibility (and
memory requirements)
• Is that because of 2D/3D encoding differences?
What happens with 3D models?
– An experiments conduced by Zack and coll. shows
that the limit comes from the transfer of
informa4on from one dimension to another (live
adult demonstra4on)
– Infants do just as well imita4ng 2D/2D than 3D/3D:
2D is not as impoverished as to block imita4on,
and 2D does not represent a poorly understood
condi4on in comparison with 3D (but live adult
demonstra4on could help the understanding)
– Representa4onal flexibility seems to be the
problem
19. Imita4on, social cogni4on & mirror
neurons
• Among the studies on social
cogni4on, mirror neurons have
gained lot of a^en4on
• Mirror neurons are involved in the
representa4on of an ac4on
• Mirror neurons are ac4vated when
observing an ac4on, independently
from the specific motor realiza4on
of the ac4on
• Mirror neurons are related to the
goal, and the agent
• Mirror neurons could be involved in
the understanding of others’
inten4ons and to imita4on
• Specula4vely, in empathy (Iacoboni,
et al., 2005)
20. Human imita4on
• Infants understand and
imitate adults’
inten4ons, not only
their behaviors
• Learning by imita4on
seems to require the
understanding of
others’ inten4ons
(Tomasello, 1990)
21. Understanding human inten4ons
• Three levels of
understanding others’
ac4ons & reading of
inten4ons)
– Perceiving others as actors
that produce their ac4ons (6
months old children)
– Perceiving others as having
goals for their ac4ons (9
months)
– Perceiving others as making
plans for reaching their goal,
and choosing the most
ra4onal ac4on (14 months)
(Tomasello, et al. 2005)
22. Engaging in shared inten4ons
• 3 levels of engagement in
shared inten4ons:
– Dyadic engagement: face
to face interac4ons and
protoconversa4ons with
shared emo4ons
– Tryadic engagement: doing
things together, but
without assigning roles for
the reaching of the goal;
sharing percep4on and
goals (9‐12 months)
– Collabora4ve engagement
= sharing ac4on plans
(12‐15 months)
23. Humanness
• At the origin of human culture
and cogni4on stand two
capaci4es:
• ‐ mind reading, and in
par4cular: the capacity of
perceiving and understanding
others’ inten4ons
• ‐ a mo4va4on for engaging in
shared inten4on ac4vi4es
• So: shared inten4onality is
what makes humans special in
the animal reign
• (Tomasello, 2005)
24. Cultural intelligence hypothesis
• Cogni4ve, Evolu4onary
anthropology
– Baby humans differ from
primates mainly because of
social abili4es
– Further differences
between humans and
primate might derive from
these social‐cultural
– Humans have developed
special cogni4ve skills as a
result of the development
of specialized skills for
absorbing knowledge and
prac4ces of their social
group
27. Natural pedagogy
• “… human communica4on is specifically • Development of natural pedagogy:
adapted to fulfil the funciton of transmiFng
generic knowledge between • Development of tools’ making prac4ces
individuals.” (Gergely & Csibra, p. 3) represents an evolu4ve pressure
• “A new type of communica4ve learning • Because these prac4ces cannot be learned/
system based on ostensive‐referen4al transmi^ed by other, available mechanisms
demonstra4ons of knowledge … expert user of learning from imita$on/observa$on*
ac4vely guide the novice by selec4vely
manifes4ng the informa4on to be acquire • Because they represent opaque contents for
and generalized. cogni4on
• … children … are always novices with respect • Thus, humans have evolved mechanisms that
to the accumulated knowledge of their serve the pedagogical func4on of
culture. transmiFng cogni4vely opaque contents
• This is why we call the specific aspects of • These mechanisms are part of the more
human communica4on that allow and general communica4on system
facilitate the transfer of generic knowledge • They consist of demonstra4on acts:
to novices Natural Pedagogy. ” (Gergely & ostensive‐referen4al demonstra4ons
Csibra, p. 4)
28. Adults/children natural pedagogical
system
• “When children are shown an ac4on • Children observe and imitate adults
performed in a par4cular style leading to a – Children spontaneously imitate causal ac4ons
clear end state (e.g. a mouse is hopping that lead to achieve goals, and ignore other
across the table into a house), they tend to components of the global ac4on
reproduce only the end state (put the mouse – The others components of the ac4on are
into the house), oken ignoring the manner of opaque to children’s cogni4on
ac4on (hopping). However, if the relevant – But, when the “teacher” makes it clear that
informa4on concerning the end state is these components of the ac4on are relevant,
communicated to them verbally by the actor children do pay a^en4on, and imitate
before the demonstra4on (“the mouse lives • Adults use their communica4on system to
in the house”), they reproduce the ac4on facilitate children’s learning
style more oken. • Young children are recep4ve to adult’s
• Ostensive communica4on does not only ostensive demonstra4on before they are able
make children pay more a^en4on to the to use it for learning
demonstra4on but they also see it as a
special opportunity to acquire generalizable
knowledge.” (Gergely & Csibra, p. 5)
• Ostensive signals allow to
• “recent studies ...demonstrate this – Disambiguate the nature of the ac4on
preparadness in the form of three kinds of (communica4on, not just using the tool)
early perceptual and cogni4ve biases: – Disambiguate the target of the
communica4on (you)
29. Ostensive signals
• Preference for ostensive • 1. preferen4al a^en4on for
signals : the sources of ostensive
– Gaze contact signals
• Newborns preferen4ally look at
schema4c face‐like pa^erns with
direct gaze vs averted gaze;
preference disappears when
faces are upside‐down;
preference disappears when the
typical iris/sclera pa^ers of eyes
is inverted
• Same neural ac4va4on for
infants and adults in response to
direct gaze and common neural
ac4va4on for two different
ostensive s4muli (direct gaze &
eye‐brow raise)
– Motherese
– Mo4onese
30. Referen4al expecta4ons
– Infants follow the gaze of interac4ng
• 2. Referen4al expecta4on
adults to iden4fy what they are looking induced by ostensive contexts
at, before they can understand language • Eight‐months olds observed
– Useful for sampling parts of someone on a computer screen
the world that others found ostensively looking at and
interes4ng, and present in gree4ng them before shiking her
other animals gaze to llok behind one of two
– Human infants followgaze barriers. Following this, an object
shiks only when these are
preceded by ostensive signals was revealed either at the
(gree4ng, gaze contact) targeted or at the other occluded
– Infants expect to find an object at the loca4on. Infants’ looking pa^ern
“end” of a gaze‐following in an ostensive suggested that they expected to
context find an object at the loca4on
– 13 months old Infants expect to where the person’s gaze wwas
find the named object (if its name directed at, just like older infants
is part of their vocabulary) do in similar live
– But not if the gesture and word are situa4ons.” (Gergely & Csibra, p.
emi^ed by different persons 5‐8)
31. Interpreta4on bias
– Not only infants are prepared to receive ostensive–referen4al • 3. interpreta4on bias to
communica4on, but they do expect to learn something generalizable
from it (and not just a par4cular instance) = to learn about referent preferen4ally encode the
kinds
– When infants (18 months old) observe adults expressing content of ostensive‐
emo4onal valence in rela4onship to an object in a non‐
communica4ve context, they infer that person’s par4cular
preference (she does not like it). But when the same pa^ern
referen4al communica4on as
of valence expression is inserted in a communica4ve
context, infants a^ach the expressed value to the object and
represen4ng generalizable
expect that other people will react in the same manner to
the object (it is disgus4ng for everybody) knowledge”
– Infants (9 months old) shik their encoding pa^ern from
loca4on to appearance features when the situa4on shiks
from non‐communica4ve to communica4ve.
• “this is what dis4nguishes our
– They are more likely to detect change in loca4on in hypothesis in the first place
a non‐communica4ve situa4on, but detect more
oken features change in a communica4ve situa4on
and neglect loca4on; and this happens even in
from compe4ng proposals,
situa4ons in which loca4on is important,
pragma4cally, such as hiding games
according to which human
– This bias could explain A not‐B task errors: children
stop being interested in loca4on and do not mind
communica4on originates
about the new loca4on, because the
communica4ve contexts has made them focus on evolu4onarily and
the features of the object. In fact, once
communica4ve cues are removed, the errors ontogene4cally from a basic
–
diminish.
Appearance features are be^er candidates for mo4ve to cooperate with
later use and object iden4fica4on, thus for
generaliza4on. others to achieve shared
– Communica4on has evolved not only for collabora4on‐purposes but goals.” (Gergely & Csibra, p.
also under the pressure of learning/teaching purposes
5‐9)
32. Social learning mechanisms
• “There are many types of social learning • Social learning mechanisms are common to several
mechanisms in the animal kingdom, and they all animal species
involve some form of observa4onal learning, where • Learning generalizable knowledge from social
the observa4on of an adap4ve behavior of another interac4ons seems to be specific to humans
individual makes it more likely that the observer will • Natural pedagogy seems to be universal, thus
produce the same or similar behaviors in the future.
In this sense, social learning represents transmission “natural”
of general knowledge or skills from one individual to
another.
• Non‐human animals communicate about episodic,
non‐generalizable informa4on (that applies only to
the here and now), and learn new skills by
observa4on or scaffolded individual learning, they
do not seem to use communica4on to pass on
generalizable knowledge to others.”
• “ This discrepancy between general claims about
the absence of teaching and the actual reports is
likely to reflect the enormous differences between
teaching in Western socie4es and in more
tradi4onal cultures. It is not just that Western
educa4on relies heavily on formal schooling but also
that it aims to provide verbal explana4on and
jus4fica4on for what is being taught. … however,
Natural Pedagogy … seems to be
universal.” (Gergely & Csibra, 2009, p. 12‐14)
33. • “Child development is today conceptualized
as an essen4ally social process, based on
incremental knowledge acquisi4on driven by
cultural experience and social context. We
have “social” brains.” (Goswami, 2008b, p. 1)
35. Distributed cogni4on
• The unit of analysis
of cogni4ve
performances
should be extended
beyond the
individual so as to
encompass social
and material
interac4ons with
tools
– (Hutchins, 1995)
36. Extended cogni4on
• Performances
typically described as
cogni4ve are
significantly worst in
absence of
interac4on with
tools, others, or of
epistemic ac4ons
that have no other
aim than favoring a
be^er knowledge of
the world
– (Clark & Chalmers,
1998)
37. Social neuroscience
• Strong accent on “… the brain does not exist in isola4on but
cogni4on as a social rather is a fundamental but interac4ng
component of a developing or aging
phenomenon which individual who is a mere actor in the larger
produces changes in theater of life. This theater is undeniably
the brain, as well as social, beginning with prenatal care,
changes in the brain mother‐infant a^achment, and early
produce social childhood experiences, and ending with
phenomena loneliness or social support and with
familiar or societal decisions about care for
• importance of the elderly. … Social psychology, with its
mul4level, panoramic focus on the effects of human
integra4ve analysis associa4on and the impact of society on
of complex the individual, is therefore a fundamental
psychological although some4mes unaknowledged
complement to the
phenomena neurosciences.” (Cacioppo & Berentson,
1992, p. 1020)
38. Integra4on of levels of analysis
importance of mul$level, integra$ve analysis of
complex psychological phenomena
1. Neurochemical events influence social processes; Social
processes influence neurochemical events
• Difficulty in the integra4on of neuroscience and social psychology
levels of analysis: different scales into which brain and behavior can
be represented
• The level of organiza4on of psychological phenomena vary from
molecular the organism set into a physical environment and a socio‐
cultural context
• Neurosciences generally encompass the lower level of the
spectrum, social psychology the higher one
• Integra4on means that analyses at each level of organiza4on can
inform, refine or constrain inferences in the other levels
39. 2. The study of the elements of the system can fall short of
useful and comprehensive explana4ons
• In other sciences, the existence of different levels of explana4on
(protons/rocks) does not lead to considering geology as a folk
theory when compared with molecular level models.
• Dis4nc4ve levels of analysis are complementary, not alterna4ve
40. – 3. A set of neural events can be a sufficient cause for
producing a psychological phenomenon, without being a
necessary one
• E.g., lying rubustly produces certain electrodermal responses ; but
other condi4ons can produce the same electrodermal responses
• E.g. schizophrenia is reliably associated with elevated dopamine
levels (elevated dopamine levels produce schizophrenia‐like
symptoms) but excessive levels of dopamine are not necessarily
involved in all cases of schizophrenia
– However, when other neurochemical mechanisms are iden4fied that
produce schizophrenia‐like symptoms with a different neurochemical
basis, it is possible to part the psychological term “schizophrenia” in
different pathologies.
• In the case of mul4ple determinants of a certain behavior, studies
on the sufficiency of a certain neurophysiological condi4on in
causing a certain phenomenological phenomenon are impôrtant
but lack generalizing power.
41. from medicine to educa4on
• “… no single level of behavioral organiza4on is best for all psychological ques4ons.
• An example can be found in the rela4ve u4lity of specifying the sociocogni4ve versus
the neurophysiological basis of pa4ent delay following the onset of gynecologic
cancer. Women can now survive most gynecologic cancers if the disease is diagnosed
and treated early. … The form of the representa4on of pa4ent delay offered by
neuroscien4fic analyses of pa4ent delay, although perhaps contribu4ng to more
complete understanding of the phenomenon, is not op4mal for iden4fying the
determinants of pa4ent delay or for developing effec4ve interven4ons to minimize
such delay. Huge savings in resources and human suffering are there to be reaped not
through a specifica4on of the brain circuits underlying pa4ent delay, but by well‐
conceived public health campaings that iden4fy the early signs of cancer…
” (Cacioppo & Berentson, 1992, p. 1022)
• “It follows … that an exclusive focus on a reduc4onis4c (e.g. neurophysiological,
molecular, gene4c) level of analysis can mask contribu4ons of other levels of
organiza4on to mental order and disorder and thereby constrain theore4cal accounts
of psychological phenomena.”
• “Hence, without a^en4on to basic social psychological factors and processes, the
decade of the brain may yield some spectacular images and experimental effects but
rather limited answers to the problems of mental health.” (Cacioppo & Berentson,
1992, p. 1025)
44. The 2 sigma problem
• Bloom, 1984 has compared 3 condi4ons of instruc4on:
– Conven4onal (1:30, periodic tests for marking)
– Mastery learning (1:30, forma4ve tests for measuring mastery & immediate
feedback)
– Tutoring (1:1 or 1:2 1:3, forma4ve tests and feedback)
• He found that the average student under tutoring was above 98% of the
students in the control class = 2 standard devia4ons above the average
of the control class
• The average student under mastery learning was about 1 standard
devia4on above the average of the control class (above 84% of the
students in the control class)
• 90% of the tutored students and 70% of the mastery learning students
a^ained levels of achievement that only 20% of the students in the
control class had achieved
– Tutoring would probably not enable the top 20% of tradi4onal instruc4on group
students to do be^er; but 80% of tradi4onal classrooms do poorly in
comparison to tutoring
– Maybe this is because teachers direct their a^en4on to some students, and
ignore others