Matt,

On 4/17/08, Matt Mahoney <[EMAIL PROTECTED]> wrote:
>
> As far as I can tell, it only gives medical advice based on your personal
> agenda.  It knows only what you program into it.


... in a way that many people can each contribute their own "advice". There
simply isn't anything out there that is currently doing this yet, leaving
only the Internet that is maybe 1% as valuable.


> > I published a paper at the first IJCNN in San Diego explaining how
> > everything pointed to wet neurons generally computing with the
> logarithms of
> > probabilities of assertions being true. That simple fact should have
> guided
> > future research, but lab researchers not being mathematicians, and
> neither
> > going to NN conferences, this guiding fact as died away like the echo of
> > some long-forgotten noise. When a tree falls in the forest...
>
> I use the same technique in my PAQ7/8 data compressors (since Dec. 2005),
> although I was not aware of your research.  A set of models independently
> estimate the probability p(0), p(1) that the next bit of input will be a 0
> or
> 1 based on past history in various contexts.  The predictions are mapped
> to x
> = log(p(1)/p(0)), combined by weighted averaging,


... thereby forming the logarithm of a weighted geometric mean,

then mapped by the inverse
> squashing function 1/(1+exp(-x)), which makes it a neural network.  Then
> the
> weights are adjusted to favor the most accurate predictions in proportion
> to
> x*(actual - predicted), a simplification of back propagation that
> minimizes
> coding cost rather than RMS prediction error.


Wouldn't be nice to have a processor that kerchunked on these at 10K/clock
cycle?

I should mention the technique works quite well.
> http://www.maximumcompression.com/data/summary_sf.php


Thanks for your comments.

Steve Richfield

-------------------------------------------
agi
Archives: http://www.listbox.com/member/archive/303/=now
RSS Feed: http://www.listbox.com/member/archive/rss/303/
Modify Your Subscription: 
http://www.listbox.com/member/?member_id=8660244&id_secret=101455710-f059c4
Powered by Listbox: http://www.listbox.com

Reply via email to