🚀 go-pugleaf

In article <1cn3lmINNe9l@oly.cs.ubc.ca> yking@cs.ubc.ca (Yossarian Yggy King) writes:
>
>I saw a talk last week by Dr. Geoff Hinton, and he mentioned that he'd
>been involved in some work (though maybe only peripherally, I'm not
>sure) on building neural nets to learn how to correct for optical
>defects in large lenses. I've probably got this wrong, but it was
>something like having the net learn the distortion parameters for each
>point on the lens, which are then corrected with a secondary lens that
>is distorted in real-time as the image is gathered to correct for the
>distortions in the primary lens.
>	He said in passing that the neural-net corrected lenses
>rivaled the resolution of the Hubble scope. Obviously (I think) you
>can only correct for distortions, not attenuations, by the atmosphere
>... once you've lost information you can't get it back.
>-- 
>Yggy King	| Thinking the world should entertain you leads to boredom
>UBC Comp Sci	| and sloth. Thinking you should entertain the world leads
>B.C., Vancouver	| to bright clothes, odd graffiti and amazing grace in
>Canada		| running for the bus.    -- Ann Herbert

One should note that neural nets are not 100% accurate on the predictions
that they can make. A good net usually has an accuracy somewhere between
80 - 90%. And when a neural network isn't correct the results can vary wildly.
This is the big disadvantage of neural nets.... noone can guarantee error
bounding. So while these new telescopes should be better than existing
telescopes I would be hesitant to compare them to Hubble. Except when it
comes to $$$$ 8).

PS: You should also be able to use neural nets to enhance Hubble images. But
    as a word of caution : Anytime you distort data to achieve results that
    you deem desirable you can lose valuable information outside your interest.


Garry Holmen