Greg Detre
Wednesday, May 14, 2003
false view
of development - linear refinement progression as you grow
why don�t you just design the end-point then?
instead: ontogenic specialisation � infants are as complex as adults,
but they flexibly reconfigure themselves over time
the right context for learning the behaviour is different the context in
which it will finally be functionally expressed
play �
looking for opportunistic learning contexts
learn with
the right tools
understanding
intentions � ability to predict/make expectations about an agents� intended
goal state
what
precedes a 4-year old�s ability to pass the false belief task
Baird &
Baldwin � adults parse action sequences along intentional lines � asked them
where the action boundaries are
hardcoded a
program to grip a beanbag
then tried
getting Ripley to learn it
went through 60 grip-trajectories
used a 14-channel 10-state HMM to recognise gripping � no false
positives
inverted
recognition to get generation is not that simple
Ripley takes seven joint positions as input, but the features are in
Cartesian coordinates
can�t simply reverse isGripping() � but you can perform the primitive
actions that co-occur with isGripping, and assert those associated states
solutions:
built-in inverse kinematics, and associative learning
could have don�t without the inverse kinematics, by getting Ripley to
motor babble etc. � no time
might be
able to get away with k-means clustering, rather than HMM � even possibly
online learning
didn�t get
a chance to try different actions
still stuck
with the salient 14 features, e.g. xyz position, distance to objects,
isGripping etc.
evidence
that interpreting body language and speech have shared underpinnings
Arbib, 2000
� �language evolved from a basic mechanism not originally related to
communication� � mirror system hypothesis