Transcript Current opinion suggests that language is a cognitive

Babies and Computers
Are They Related? – Abel Nyamapfene
Abstract:
Current opinion suggests that language is a cognitive process in
which different modalities such as perceptual entities,
communicative intentions and speech are inextricably linked. In
this talk I discuss my belief that the problems psychologists are
grappling with in child development are also the same problems
computer scientists working in artificial intelligence and robotics
are facing. I show how computational modelling, in conjunction
with the availability of empirical data, has contributed to our
understanding of child language acquisition, and how this
knowledge has advanced progress in robotics.
Psychologist
Computer Scientist
How do babies
learn life skills?
How can you be
as adaptive as a
baby?
Basic Computer Organisation
Von Neumann Architecture
• stored program: data and
programs are stored together
• sequential control:
programs that are executed
sequentially.
• Algorithmic: Everything to
be done defined beforehand
• Program implements
algorithm in computer
friendly language
Von Neumann Architecture Pros & Cons
Good for procedures
that can be
pre-defined before
execution: e.g:
• numerical
computation
• Word processing
• Car assembly
• Precision surgery
Poor for procedures that
have to bee adapted on a
situation by situation
basis e.g:
• Language processing
• Pattern processing
• Artificial human
assistant
Emerging Computer Applications
• Social Interaction
– caregivers
– domestic
– helpmates
•
•
•
•
Intelligent weaponry
Games
Medicine
Education
Examples
Games
humanoids
Education
Weapons of War
Medical
Diagnostics
Features Common To Intelligent
Computer Applications
• Computer applications still fall far short of
expectations
• Applications only work well within well specified
environments
• Application scalability is limited
• Processing capability has little or no incremental
capability
In Comparison:
Children come into the world with little or no cognitive
skills but exhibit developmental progression of
increasing processing power and complexity. An
example is language where children progress from no
language, to babbling, to one-word utterances, twoword utterances and finally full adult speech – almost
all the children .
What can Computing learn from Children?
Learning from Child Development
1: Carry out Empirical Investigations of
Developmental Activities
- Behavioural Investigation
- Neuroscientific Investigation
2: Use Empirical Data to develop Models of
Development process
3:Assess and Incrementally Improve the Models
4:Apply knowledge to computer tasks
Empirical Investigation:
Behavioural
• Observe developmental activity – e.g. language
acquisition
– Track single child from conception to stage of
full acquisition – “Keep a Diary”
– Study sizeable number of children at same
stage of development
– Carry out ethically approved psychological
investigations on children etc
Empirical Investigation:
Neuroscientific
Investigate:
• Brain Maturation
Processes
• Interaction of Brain
Regions
• Interaction of
Individual Neurons
Models of Development Based on Brain Neural
Processing
Actual Neurons: Complex
Models of Development Based on Brain Neural
Processing
Artificial Neurons: Very Very Simplified
Some Models of One-Word Child Language
“Dada” instead of “Here comes Daddy.”
“Uh oh” instead of “I am happy.”
“More” instead of “Give me some more”
1: A multilayer perceptron network for mapping
images to text (Plunkett et al, 1992).
Image (output)
Label (output)
joint internal
representation
Label
representation
Image
representation
Image (input)
Label (input)
Network by Plunkett et al simulates word – image association and exhibits
same developmental learning as a child, but learning mechanism not
biologically feasible
2: Hebbian-linked Self –Organising Architecture
Li, Farkas & MacWhinney (2004)
Perceptual Input
Unidirectional links from Perception
to Speech Neuron Layers
Second SOM
First SOM
Unidirectional links from Speech and
Perception Neuron Layers
activated neuron
Speech Input
Network was inspired by the belief that Brain Modules are interlinked. It
successfully simulates Word-Object Mapping in children
3: An Approach that can associate Two Input Types: Full counterpropagation network
(Hecht-Nielsen,1987)
x input
layer
Z1
y input
layer
Z2
x output
layer
ZN
cluster
layer
y output
layer
4: Extending the Counterpropagation Approach to
Modelling Child Language
(Nyamapfene &Ahmad, 2007)
Competitive Neuron
layer
Modal
weights
Perceptual Input
Speech Input
Intentional
Input
Model based on empirical evidence that children have
intentions and that brain has multimodal neurons
I have described some investigations of child
language acquisition through:
• Physically observing infants acquiring language
• Studying relevant brain structures
• Building, testing and modifying brain inspired
computer models of child language acquisition.
Current Conclusions on Child Language
Acquisition Suggest That:
• Child language has multiple inputs that need to be
processed simultaneously
• Language acquisition takes place through social
interaction with caregivers
• Children have desires, have emotions, set and modify
goals, monitor ongoing speech acts and generate
communicative intentions which lead to speech
utterances
5: A Control-Theoretic Neural Multi-Net Model of
Child Language Acquisition
(Nyamapfene, 2008)
Child
Environment
Desires
Emotions
Drive
Goals
Communicative
intentions
Single-Word
Utterance
Caregiver
response
Block diagram of a control systems approach to modelling child language
at the one-word early child language acquisition stage
From Child Development To Computing
Cynthia Breazeal has
developed Kismet, a
robot that employs drives
and emotions to interact
with a human – based
on social interaction of
an infant and a caregiver
(Breazeal and
Brooks, 2004)
Current & Future Projects
• Developing a multimodal neural network model
that learns from Child - directed Speech using
cross-situational techniques
• Implementing the control-theoretic model of child
language acquisition presented in this talk using
neural multi-nets
• Migrating child work onto a robotic platform –
(circa 2009 – 2010)
Finally: Yes, I Think Babies and Computers are Related
Thank You!!??!!
References
•
C. Breazeal and R. Brooks (2004). "Robot Emotion: A Functional
Perspective," In J.-M. Fellous and M. Arbib (eds.) Who Needs Emotions: The
Brain Meets the Robot, MIT Press (forthcoming 2004).
•
R. Hecht-Nielsen (1987). “Counterpropagation Networks,” Applied Optics
26:4979-4984.
•
P. Li, I. Farkas, B. MacWhinney (2004). “Early lexical development in a selforganizing neural network,” Neural Networks 17: 1345 - 1362
•
A. Nyamapfene (2008). “Computational Investigation of Early Child
Language Acquisition Using Multimodal Neural Networks: A Review of Three
Models,” Artificial Intelligence Review (submitted).
•
A. Nyamapfene and K. Ahmad (2007). “A Multimodal Model of Child
Language Acquisition at the One-Word Stage,” 20th IJCNN: International
Joint Conference on Neural Networks, 12th-17th August, 2007, Orlando,
Florida, USA
•
K. Plunkett , C. Sinha, MF. Muller, O. Strandsby (1992). “Symbol grounding
or the emergence of s
symbols? Vocabulary growth in children and a
connectionist net,” Connection Science 4: 293-312