[From Rick Marken (2012.10.08.1800)]
Sigh!
RKM: Sorry to disappoint Warren, but when you say that you agree with Rick, do
you mean that that you agree ...
- it's OK to voice an opinion on a subject (i.e. the content of Wolpert's
talk) without fully reviewing the material in question?
RM: I'm really sorry if my reply upset you Roger. I should have been
much more diplomatic. It's true that I didn't watch the entire TED
talk but I don't think it was necessary. You had posted the video
saying that there's lots of PCT-relevant arguments; but the whole
thing was about controlling actions (through Bayesian inference,
predictive control, action selection etc) and control of action is not
a part of PCT, nor has there been any data produced to show that it
needs to be a part of PCT. There may be situations where prediction is
necessary -- I think Powers has a model of eye movement that requires
some prediction -- but even when prediction is used it's part of a
closed loop control process that fixes things up after the
prediction-based movement is made.
But the big problem with Wolpert's talk, from my perspective, is that
he starts with the assumption that it is movement itself that is
controlled -- which is precisely wrong from the PCT perspective;
prediction or no prediction, what is controlled has to be a perceived
consequence of movement. Based on this assumption (control of output)
Wolpert concludes that prediction is always involved in the production
of movements.
PCT, on the other hand, starts with the assumption that it is the
perceptual consequences of movements -- not the movements themselves -
that are controlled. Models based on this assumption account for a
lot of movement data highly accurately without any need for the
incorporation of prediction. So PCT leads us to assume that prediction
is not involved in control until we are shown otherwise.
RKM: - it's acceptable to publicly question the veracity of someone's
scientific reputation without providing specific reasons?
RM: I don't believe I ever questioned the veracity of Wolpert's
scientific reputation. I was questioning the relevance of his work to
PCT. From a PCT perspective, Wolpert is on exactly the wrong track.
That doesn't mean he's a bad scientist; just that he doesn't seem to
have read any PCT and if he has he hasn't understood or been convinced
by what he's read. Heck, according to Einstein, the greatest
physicists of his time were wrong but that didn't mean that Einstein
was questioning their scientific reputations.
RKM: - it's productive to express satisfaction that a related field of study
seems not to have all the answers to a bunch of seriously hard problems that
face us all?
RM: Huh? To quote my friend Bob "No I don't feel so good when I see
the heartaches you embrace". I take no satisfaction when I see an
obviously very smart researcher headed down what I know is the wrong
path. But PCT has been rather harshly rejected by the "control of
movement" people (for what are clearly political rather than
scientific reasons; PCT has never been rejected by data) so it's a
little tiresome to have another "predictive control of output"
theorist trotted out as someone we could learn something from.
RKM: - it's correct to claim that "PCT has shown that all of the arguments are
incorrect" (i.e. Bayesian inference, predictive control, algorithms for
separating 'self' from 'other', action selection etc.) without citing the
relevant peer-reviewed evidence to back it up in each case?
RM: I didn't know that your post was supposed to be the start of such
a discussion. Predictive control has been a topic on CSGNet many times
but I would be happy to have that discussion again if you like. I can
try to point to relevant research and references once we get started.
RKM: I had hoped that people would see the potential overlaps and opportunities
to move our science and understanding forward collectively.
RM: I guess your hope (prediction) didn't pan out that well; that's
the problem with prediction; it's not nearly as reliable as control of
input;-) Perhaps it would have helped if you had used Bayesian
prediction. Then you could have included in your calculation the fact
that among those of us who do research on PCT the prior probability of
seeing overlap between a control of output model and PCT is extremely
low. But I am certainly willing to discuss this; if you can show me
some data that seems to demand Bayesian inference, predictive control
or action selection, then I could be convinced that there is overlap
between PCT and what Wolpert was talking about. So let's have the
discussion.
Best
Rick
···
On Mon, Oct 8, 2012 at 10:30 AM, R K Moore <r.k.moore@dcs.shef.ac.uk> wrote:
---
Richard S. Marken PhD
rsmarken@gmail.com
www.mindreadings.com