Greg Detre
Tuesday, April 22, 2003
Schaal
emphasises the importance of research into imitation learning as �[channeling]
investigations in computational motor control towards the important topic of
action-perception coupling�. Imitation (in its various stronger and weaker
forms) seems to play a central role in learning perceptual-motor coupling by
directing the search towards valuable actions in a given situation, which would
give rise to learning at a much much faster rate than trial and error or
statistical approaches. Or, as Schaal puts it, the only way to search huge
state-action spaces is to either find more compact state-action representations
or to know which bits of the space are most relevant. The imitation learning
approaches he discusses can help with both.
It is with
regard to the first question that he discusses �movement primitives�. I found
this to be a misnomer, since the movement primitives he describes are by no means
primitive or atomic, as one might expect (as in the animation of Dobie).
Basically, the idea seems to be to break the total action-sequence to be
described into a sequence of relatively stereotyped or common low-level
actions, and then to encode them as an aggregate in the service of a goal. This
way, the issue of combinatorial explosion is avoided, since movement primitives
are compact and can be tailored with parameters to apply in multiple domains
and slightly different situations. High-level movement primitives could rely on
the via-point method, splining or feedforward models to generate the low-level
commands and to move through arbitrary trajectories.
This rests
on one of the most important ideas, that of �movement recognition � based on
the movement generation system�. That is, the system starts by building an
internal feedforward model of its own actions to try and predict what will
happen when a given motor command is issued, using supervised learning. This
feedforward model can then be employed when trying to imitate, to internally
test which motor command will produce the desired result. When you have
multiple feedforward models, which might different in the representations they
employ, the limbs they emphasise or trained on different actions in the past,
these can compete so that the system can choose which appears to offer the most
reliable prediction in a given situation.
I wasn�t
entirely happy with the distinction made between task-strategy and task-goal.
This seems to be important, since he defines true imitation as being present
only if:
1. the imitated behavior is new for the
imitator
2. the same task strategy as that of
the demonstrator is employed
3. the same task goal is accomplished
I didn�t
see a definition of either �task strategy� or �task goal�. They might be
analogous to the difference between action-level imitation (�the indiscriminate
copying of the actions of the teacher without mapping them onto more abstract
motor representation) and program-level imitation (�a process by which the structural
organization of a behavior is copied from observing a teacher, while the exact
details of actions are filled in by individual learning�). In other words, the
task strategy would be some medium-level description of how to achieve the
task, and the task goal would be a goal-state or abstract description of why
the task is being performed. Unfortunately though, this distinction is
problematic. After all, presumably, the task strategies we employ in later life
are composed out of the simple task goals we learned in the past, implying that
there is some commonality of representation, and that they might be collapsed
into a single hierarchy. Another problem involves the author�s decision to
stick to visually-mediated imitation, since language seems necessary in most
situations to communicate and share a goal in the first place.
Finally,
the notion that the goals of the student and of the teacher can be declared
unproblematically to be the �same� masks a necessary mapping that has to be
made between them. When imitating a tennis swing, I may have to map the goal of
�his hitting that ball with his right arm holding that tennis racket� to �me
hitting this ball with my left arm (perhaps) holding this tennis racket�. This
is the �correspondence� problem, of mapping coordinate frames. But the problem
may be deeper � I may be supposed to be imitating the direction of shot, or the
facial expression or indeed anything about the situation. There needs to be
some means for top-down goal-like knowledge to inform what about the teachers�
lower-level actions should be considered salient. This issue of �what about the
situation to imitate� is reflected in the discussion of a cost function J,
which needs to capture both the task goal and the quality of imitation in
achieving the task goal.
Schaal�s
discussion of imitation (especially the neuroscientific evidence) almost seems
to assume that there is only one motor-imitation system in the brain. We know
that mirror neurons are specific to special motor behaviours with a particular
object, as executed by physiologically �similar� beings. As a result, it�s
conceivable that (at least in humans) there�s some other, more general
imitative system that we use when imitating non-humans, or imitating arbitrary
actions with uncommon limb movements. There may be more fine-grained domains of
specificity for mirror neurons, e.g. for voices, facial expressions. Perhaps
signature-forgers have mirror neurons that respond to certain kinds of
hand-writing movements. There may, as Schaal seems to believe, be some division
between imitation of a goal and imitation of a motor performance.