hibbert at mydruthers.com
Tue Dec 18 04:07:50 UTC 2007
>>> The place I'd start is with a very careful read of Jeff Hawkins'
>>> "On Intelligence".
And Eugen replied:
> Sorry, the book is not even wrong.
This doesn't give me enough information to tell whether to believe you
or what I read earlier. What's the simple demonstration that it's "not
even wrong"? Is there a better model for how intelligence emerges from
collection of neurons or for how the cortical columns work together to
produce interesting behavior?
The book starts by arguing that if we are to understand how the brain
works, we need to realize that the cortex consists of mostly uniform
components at the level of the columns, that they usually perform
functions determined by where they are located in the brain, but that
when there's damage, any particular area might take on a function
performed elsewhere. Then he looks at what is know of the structure of
the columns and comes up with an architecture that, IMHO, would work as
a matter of remembering inputs, predicting their recurrence, and feeding
the predictions back to the earlier stages to serve as a gauge of surprise.
> His stuff is a) not new and b) so far can't even reproduce hoary old
> PDP models.
The claim that it's not new isn't relevant if the theories are clearly,
demonstrably wrong. If the theory is useful, it may be a valid
criticism of the author or the book, but not of the theory. I read lots
of books that present material discovered or developed by someone else.
They don't always give appropriate credit, but sometimes turn out to be
the primary citation on the subject later. I prefer authors who give
appropriate credit, but if they explain a subject well, I value them
even if they don't give credit.
Why is it important that it be able to reproduce the "hoary old PDP
models"? (I assume you mean neural nets? I found more than one
expansion in the context of neurophysiology.)
>>> A simulation at this level would be a lot cheaper, and if Hawkins
>>> is right, would capture the essential emergent properties of the
>>> cortex' building blocks.
> It would be nice if there was something to be right on. Unfortunately,
> there isn't.
What do you mean? There's no useful level to simulate between neurons
and brains? I'm guessing, and not finding any plausible expansion for
"there isn't anything to be right on".
> What irks me most is that there's this pall of standstill spreading
> over multiple fields, and for a long while now. That might be purely
> subjective, but I'm afraid it goes beyond that.
I can't tell what you're trying to say here.
* "Hawkins didn't add anything new to the field, nor has anyone else in
* "Everyone thinks there's no progress here, but they're missing the
* something else entirely
It is easy to turn an aquarium into fish soup, but not so
easy to turn fish soup back into an aquarium.
-- Lech Walesa on reverting to a market economy.
hibbert at mydruthers.com
More information about the extropy-chat