Infant Learning TEX 2016 (Language Learning), SISSA, Trieste Scott - - PowerPoint PPT Presentation

Scott P. Johnson Department of Psychology Department of Psychiatry and Biobehavioral Sciences UCLA Los Angeles, California, USA

Infant Learning TEX 2016 (Language Learning), SISSA, Trieste

same or different?

same or different?

statistical learning in infancy

familiarization

...golabutupirobidakupadotitupiropadotigolabubidakupadotitupirobidakugolabutupiropadotibidakutupirogolabubidakupadotitupirobidakugolabu... ...golabutupirobidakupadotitupiropadotigolabubidakupadotitupirobidakugolabutupiropadotibidakutupirogolabubidakupadotitupirobidakugolabu...

...golabutupirobidakupadotitupiropadotigolabubidakupadotitupirobidakugolabutupiropadotibidakutupirogolabubidakupadotitupirobidakugolabu...

test 2: ...ropado ropado ropado... test 1: ...tupiro tupiro tupiro...

...tupiro... ...ropado...

transitional probabilities (TPs): .33 1.0 transitional probabilities (TPs): 1.0 1.0

words vs. nonwords words vs. part-words

(auditory) statistical learning in infancy

• Neonates: SL of syllables in speech (Teinonen et al., 2009)
• 6-8 mo: learning of statistical distribution of phonemes (Maye et al., 2002)
• 8 mo: SL of speech sounds in an artificial language (Saffran et al., 1996)
• 8 mo: SL stems from sensitivity to TPs, not frequency (Aslin et al., 1998)
• 8 mo: SL in tone sequences (Saffran et al., 1998)
• 8 mo: SL in an unfamiliar natural language (Pelucchi et al., 2009a)
• 8 mo: SL of backward TPs (Pelucchi et al., 2009b)
• 8 mo: outputs of SL serve as candidate words embedded in new contexts (Saffran, 2001)
• 10 mo: SL constrained by sensitivity to distribution of word length (Lew-Williams & Saffran, 2012)
• 12 mo: SL of adjacent & nonadjacent dependencies (Lany & Gomez, 2008)
• 17 mo: SL of word associations in high- but not low-TP settings (Hay et al., 2011)
• 17 mo: outputs of SL serve as candidate words, associated with novel objects (Graf Estes et al., 2007)
• 22 mo: outputs of SL serve as candidate words, associated with object categories (Lany & Saffran, 2010)
• etc.

rule learning in infancy

Marcus, Vijayan, Bandi Rao, & Vishton (1999)

familiarization

Marcus, Vijayan, Bandi Rao, & Vishton (1999)

familiar test

familiar or novel?

Marcus, Vijayan, Bandi Rao, & Vishton (1999)

novel test

familiar or novel?

Marcus, Vijayan, Bandi Rao, & Vishton (1999)

2 4 6 8 10

ABA vs. ABB; ABB vs. ABA ABB vs. AAB; AAB vs. ABB Looking times (s) Familiar Novel Marcus, Vijayan, Bandi Rao, & Vishton (1999)

visual statistical learning (VSL) in infancy

visual statistical learning (VSL) in infancy

structured

{ { { { {

time 1.0 1.0 1.0 .33 .33 TPs:

Habituation Trials 1) 47.6 seconds… 5) 13.9 seconds 6) 6.7 seconds 7) 3.6 seconds Test Trials 1) Structured: 4.6 seconds 2) Random: 14.2 seconds 3) Structured: 4.4 seconds 4) Random…

“These results are consistent with the existence

• f a domain general statistical learning device

that is available to even very young infants (indeed, we found no evidence that the older infants were better able to compute the statistical structure in the input than the youngest infants we tested). Given the youngest age tested in addition to the lack of

• bserved development, it seems reasonable to

posit an associative mechanism that is functional with the onset of visual experience.” (Kirkham et al., 2002, p. B40)

visual statistical learning in infancy

“These results are consistent with the existence

• f a domain general statistical learning device

that is available to even very young infants (indeed, we found no evidence that the older infants were better able to compute the statistical structure in the input than the youngest infants we tested). Given the youngest age tested in addition to the lack of

• bserved development, it seems reasonable to

posit an associative mechanism that is functional with the onset of visual experience.” (Kirkham et al., 2002, p. B40)

• f a domain general statistical learning device

functional with the onset of visual experience.”

visual statistical learning in infancy

visual statistical learning in infancy

“These results are consistent with the existence

• f a domain general statistical learning device

that is available to even very young infants (indeed, we found no evidence that the older infants were better able to compute the statistical structure in the input than the youngest infants we tested). Given the youngest age tested in addition to the lack of

• bserved development, it seems reasonable to

posit an associative mechanism that is functional with the onset of visual experience.” (Kirkham et al., 2002, p. B40)

• f a domain general statistical learning device

functional with the onset of visual experience.”

visual statistical learning in infancy

visual rule learning in infancy

2 4 6 8 10

ABA vs. ABB; ABB vs. ABA ABB vs. AAB; AAB vs. ABB Looking times (s) Familiar Novel

habituation: three triplets

time time

ABB ABA

time

test: three triplets

visual rule learning in infancy

visual rule learning in infancy

time

habituation: four triplets test familiar 3-1-2 TPs 1.0 - 1.0 .33 - 1.0

time time

visual statistical learning in infancy

time

habituation: four triplets test familiar 3-1-2 TPs 1.0 - 1.0 .33 - 1.0

time time novelty preference

visual statistical learning in infancy

visual statistical learning in infancy

habituation: three pairs test TPs

time

familiar 1.0 2-1 .33

time time

visual statistical learning in infancy

habituation: three pairs test TPs

time

familiar 1.0 2-1 .33

time time novelty preference

visual statistical learning in infancy

!#!!" !#$!" !#%!" !#&!" !#'!" !#(!" !#)!" !#*!" !#+!" !#,!" $#!!"

!!"#$%&%#'!())))!*'+,+'-+.!))))/!0)1!

visual statistical learning in infancy

A compendium of findings:

• VSL is facilitated by familiar stimuli (faces)

time time

What are the constraints on statistical learning in infancy?

A compendium of findings:

• VSL is facilitated by familiar stimuli (faces)
• Visual rule learning is facilitated by familiar stimuli (faces)

time time

What are the constraints on statistical learning in infancy?

A compendium of findings:

• VSL is facilitated by familiar stimuli (faces)
• Visual rule learning is facilitated by familiar stimuli (faces)
• Auditory rule learning is facilitated by familiar stimuli (speech)

What are the constraints on statistical learning in infancy?

A compendium of findings:

• VSL is facilitated by familiar stimuli (faces)
• Visual rule learning is facilitated by familiar stimuli (faces)
• Auditory rule learning is facilitated by familiar stimuli (speech)
• Visual rule learning is not facilitated by another communicative signal (sign language)

What are the constraints on statistical learning in infancy?

A compendium of findings:

• VSL is facilitated by familiar stimuli (faces)
• Visual rule learning is facilitated by familiar stimuli (faces)
• Auditory rule learning is facilitated by familiar stimuli (speech)
• Visual rule learning is not facilitated by another communicative signal (sign language)
• Visual rule learning is facilitated by intermodal presentation (cf. Thiessen, 2012)

time

What are the constraints on statistical learning in infancy?

What are the mechanisms of statistical learning in infancy?

local redundancy predicts the end of a trial

“...infants were not learning global properties of the grammars presented, but were responding to local redundancy over a six-item window” (p. 1142).

• Expt. 1: pairs vs. random
• Expt. 2: pairs vs. triplets

What are the mechanisms of statistical learning in infancy?

Identification of similarity and redundancy are the core of numerous theories of learning:

• Aristotle: law of similarity
• Hume: law of resemblance
• Berkeley: likeness principle
• Ebbinghaus: law of association (association by contiguity)
• Gestalt principle of similarity (Wertheimer, 1938)
• Universal law of generalization for psychological science (Shepard, 1987)
• ACCESS model (Align Candidates, Compare, Evaluate Statistical/Social Significance) of pattern learning

(Goldstein et al., 2010)

• Intersensory Redundancy Hypothesis (Bahrick, Lickliter, & Flom, 2004)
• Repetition as a perceptual primitive (Endress et al., 2005)
• Statistical computations of similarity facilitate learning nonadjacent relations (Newport & Aslin, 2004)

What are the mechanisms of statistical learning in infancy?

habituation: three triplets

time time

ABB ABA

time

test: three triplets

What are the mechanisms of statistical learning in infancy?

What are the mechanisms of statistical learning in infancy?

• 1a. ABCDEFF
• r
• 1b. ABCDDEF

familiarization (1a or 1b) (...evidence for) “a perceptual primitive for repetition detection that is constrained by the perceptual salience of the various repetition positions within a sequence” (Endress et al., 2005, p. 441) “Salient perceptual dimensions...constrain the statistical patterns that learners most readily

• acquire. Temporal proximity is a powerful

constraint: Learners rapidly acquire the statistical patterns among elements that immediately follow each other” (Aslin & Newport, 2012)

What are the mechanisms of statistical learning in infancy?

Gary Marcus NYU

Expt.1: learning adjacent identity anywhere Habituation: ABBC DDEF GHII DEEF AABC DEFF GGHI etc. Test: JKKL MMNO PQRR MNNO JKLL etc. vs. pseudorandom (ONRQ PLNK JNPQ NQOK etc.)

• Expt. 2: learning medial adjacent identity

Habituation: ABBC ADDC AFFC ADDC ABBC AFFC etc. Test: JGGK LHHM NIIO LHHN etc. vs. pseudorandom

• Expt. 3: learning nonadjacent dependency

Habituation: ABBC ADDC AFFC ADDC ABBC AFFC etc. Test: AGHC AIJC AKLC AGHC etc. vs. pseudorandom 11-mo

p < .05

11-mo 14-mo

<

p < .05 p < .05

11-mo 14-mo

<

p < .05 p < .05

What are the mechanisms of statistical learning in infancy?

Can infants discriminate abstract “identity” from “difference” relations?

What are the mechanisms of statistical learning in infancy?

Are abstract “identity” relations more salient than “difference” relations?

!" !#$" !#%" !#&" !#'" !#(" !#)" !#*" !#+" !#," $"

!"#$#"%#&'#(')##*+&,'

!"#$%&'( !"#$%$&'$(

3/15 prefer identity p < .05

What are the mechanisms of statistical learning in infancy?

So, do infants learn better from complex or redundant information?

What are the mechanisms of statistical learning in infancy?

One learning mechanism, or two?

time time

possible strings

SVDRFV SVRFVDX FVDRFVDX

One learning mechanism, or two?

SVDRFV SVRFVDX FVDRFVDX possible chunks SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX

better chunks

SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX SVDRFV SVRFVDX FVDRFVDX

One learning mechanism, or two?

• ut

in

• ut

in

• ut

One learning mechanism, or two?

5 10 15 20 25 Grammatical Ungrammatical Mean Looking Time (s) Test Type

5- and 12-mo-olds

time

Exposure phase: 60 shape pairs 500 ms per item 500 - 750 ms ISI

time

Test phase: 200 shapes max 500 ms per item 500 - 750 ms ISI 90% expected, 10% unexpected

expected expected unexpected expected unexpected expected

Why we think understanding statistical learning is important

greater variability in the ASD group

Why we think understanding statistical learning is important

N1: visual discrimination P300: novelty detection ** p < .05; * p < .10

Why we think understanding statistical learning is important

ASD sample (n = 45)

p = .06 p = .019

Why we think understanding statistical learning is important

Why we think understanding statistical learning is important

Why we think understanding statistical learning is important

time

habituation: three triplets test “word” “part-word” TPs 1.0 - 1.0 .33 - 1.0

time time novelty preference

Why we think understanding statistical learning is important

!" #" $" %" &" '!" '#" '$" '%" '&" !(#" !()" !()" !($" !($" !(*" !(*" !(%" !(%" !(+" !"#$%&'()*%+,-%.(&/% 0(&123451637(8%8(9*:1;%<)*=*)*8'*%+>%.(&?%8%@%AA/% ,-./" 0123145"

• 67831314"

96:;<=7=5"

• 67;>=?3@A"

references to objects, people, and events not present plural, possessive, progressive, and regular past tense combining words M length of morphemes in longest 3 sentences sentence complexity

conclusions

what have we learned (so far)?

• little evidence for domain-generality of SL
• detection of TPs in visual sequences is a challenge
• VSL performance has some relation to cognitive & social function

what remains to be learned?

• how does VSL work? TPs? counting? statistical computations? chunking?

auditory chunking = “words,” “phrases,” “sentences” visual chunking = “groups,” “arrays,” “objects” what is visual sequential chunking? = “lists?” “streams?” “rapid scene perception?”

• how well can infants process & retain sequential information?
• what is the role of SL in cognitive development more broadly?
• do impairments in SL support a role for a domain-general view of ASD?
• can VSL and RL tasks be used to identify the specific hypotheses infants formed?
• what are the priors? where do they come from?
• how many learning mechanisms?