Kingma, D.P. wrote:
On Mon, Mar 3, 2008 at 6:39 PM, Richard Loosemore <[EMAIL PROTECTED]
<mailto:[EMAIL PROTECTED]>> wrote:
The problems with bolting together NN and GA are so numerous it is hard
to know where to begin. For one thing, you cannot represent structured
information with NNs unless you go to some trouble to add extra
architecture. Most NNs can only cope with single concepts learned in
isolation, so if you show a visual field containing 5,000 copies of the
letter 'A', all that happens is that the 'A' neuron fires.
If you do find some way to get around this problem, your solution will
end up being the tail that wags the dog: the NN itself will fade into
relative insignificance compared to your solution.
Well, you could achieve that (5000 registration of the letter 'A' with
their corresponding position in the image) by using a sliding window
over multiple rescaled (and maybe other transformations) transformations
of the input image. This way, you get image patches for each window and
scale (and maybe other transformations), and each patch can be a given a
corresponding position in multidimensional space (e.g., an image patch
with X and Y position and scale S has is a point in 3-dimensional
space). For each of the produced points (patches) in the space, run the
neural net to produce a lower-dimensional code and corresponding energy
(= reconstruction quality). Now filter this space by let the points have
local battles for salience using some heuristic (e.g. lower energy means
higher salience) and filter out the low-salient points. This produces a
filtered space with fewer points then the previous one, and each point
containing a lower-dimensional code.
In the example of the letter 'A', the above method would recognize all
5000 versions while remembering their individual input position. This
presumes the neural net is properly trained on the letter 'A' and can
properly reconstuct them (using Hinton's method). This should produce
5000 registrations of the letter 'A', while filtering out unimportant
information.
But you could take it a step further. For each image input, the above
method creates a filtered, 3-dimensional space with points containing
low-dimensional codes. This space can then again be harvested by taking
patches with each patch containing /n/ points, each point containing an
/m /dimensional code, so each patch being (/m/*/n/)./ /A neural net can
be trained on lowering the dimension of these patches from (/m/*/n/) to
something lower-dimensional. This process is quite similar to the one in
the previous paragraph.
What could /possibly /go wrong? :)
Regards,
Durk Kingma
Excellent! Sounds like a perfect solution ;-).
Oh, wait!
What about......... if the scene is structured in such a way that the
5,000 copies of the letter 'A' were actually scattered around in such a
way that most (but not all) of them were arranged to form a huge letter 'A'?
Would it then count 5,001 copies?
Oh, and one more thing I forgot to mention that is in the same scene
(how could I forget this one?): there are also a couple of women
standing side by side, leaning against each other with their shoulders
touching and keeping their bodies stiff and straight, forming the two
sides of a letter 'A', and holding a model of a horizontally reclining
woman between them at waist height, to form the crossbar of a letter 'A'.
Could we get the NN to recognize, in the context of the overall scene,
that here were actually 5,002 copies of the letter 'A'......?
And if the scene had one single, rather small letter B over in the
corner, would the NN find this funny?
You have 30 minutes to devise an algorithm, Durk... :-).
Richard Loosemore
-------------------------------------------
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=95818715-a78a9b
Powered by Listbox: http://www.listbox.com
