|
GatKong |
|
Tetris Mason
|
|
|
|
|
|
Reged: 04/20/07
|
|
Posts: 5907
|
|
Loc: Sector 9
|
|
|
|
Send PM
|
|
Re: Neat try, but not really comparable.
04/23/13 08:38 PM
|
|
|
May I ramble on the learning part?
Whilst in the Army, we were tryign to design software that could find tanks hidden in the bushes, and decern friendly from enemy... ultimately of course to destroy enemy ones with a robotic tank killing machine. We used photographs to train the computer.
My point is that you can't always control what data points the computer is using for its value judgements. This guys game learned to increase scores etc, and may have found "exploits", but here's what happened to us.
We were 100% on all our test photographs at identifying tanks and labeling only enemy ones.
Woot.
When it came time to demonstrate this live for the top brass... the stupid fucking thing found and "destroyed" every single tank without discriminating enemy from friendly, and the whole thing was scrapped as worthless.
Only after looking at it later why it failed so miserably did we figure out....
All the photos of the enemy tanks were taken at the beginning of the day... and all the friendly tanks were taken at the end of the day...
So the computer DID learn to discern the tanks... by the position of the sun! And we did the demonstration in the early day... thus, every tank was judged as enemy.
In retrospect, the program was a success... but alas the funding was gone and it was dead int he water.
Personally, I found the guy's video awesome. He's helpign the computer focus on the important by eliminating distracting variables like visual outputs and music, etc. Baby steps. 
Imagine someday a computer than can detect a liar because of a "tell" which people ignore, but which the computer discovered and exploits.
All good.
|
|
Index 
Flat 
Re: Neat try, but not really comparable.
Reply
