[Sussex] partitions and the like ??

[Steve Dobson]

Geoff

On Mon, Jan 03, 2005 at 11:03:31PM +0000, Geoffrey Teale wrote:
> Which reminds me of one of my favourite koans:
> 
> In the days when Sussman was a novice, Minsky once came to him as he sat 
> hacking at the PDP-6.
> 
> "What are you doing?", asked Minsky.
> 
> "I am training a randomly wired neural net to play Tic-tac-toe" Sussman 
> replied.
> 
> "Why is the net wired randomly?", asked Minsky.
> 
> "I do not want it to have any preconceptions of how to play", Sussman said.
> 
> Minsky then shut his eyes.
> 
> "Why do you close your eyes?", Sussman asked his teacher.
> 
> "So that the room will be empty."
> 
> At that moment, Sussman was enlightened.

Did he manage to train the neural net?  They are notoriously difficult
to train correctly.

I was told this story that the British Army once decided to build a neural
net to identify tanks on the battle field.  Today vast amounts of imagery
(and other data) are gathered & sifting thought this to find the one or
two percent of pictures that are worth in depth study is a large task.

So off they went to Sailsibury plane (or wherever) and took loads of photos
with tanks in them.  Sometimes the tanks were on the move, sometimes they
here partly/mostly hidden by trees or bushes.  But a vast array of photos
with various parts, of tanks seen from all angles.

With a neural net you don't just need positive samples, you also need 
negative samples too. So a few days later they when back and took loads
more photos of the same place when the tanks were not there.

Then then trained the system.  They would input a photo with a tank in it
and tell the system that this photo contains a tank, but not where it was or
how much of it was showing.  Or they would input a photo and tell the system
that it contain no tank at all.  The system had to "figure out" for itself
what a tank looked like.  They kept some photos back, so they could test
the system, and they did get it to the point where they could input a photo
and the system would tell them if there was a tank present or not.

Then the big wigs came down, and brought with them a new set of photos. 
The system failed.  It would even report a tank in a photo of an empty 
field.  The problem with a neural net is that you can't look at the links
and tell what it has learnt.  So they had to have a close look at the photos
and try and figure out what they had taught the system.  It was only then
that someone notices that all the pictures with tanks were taken on a 
cloudless day.  They had trained the system to tell them if the sun was 
out or not.

Steve