|
Re: Neat try, but not really comparable.
04/30/13 03:23 PM
|
|
|
> Also, no offense, but I find your Army story a little implausable. Surely anyone > implementing such a system would do some trial runs before demoing the thing. And > surely you wouldn't just happen to have every trial run use the same photos of tanks > as the others, and surely the additional photos of tanks used in the trial runs > wouldn't just also happen to have enemy tanks always photoed later in the day than > friendly tanks.
Never heard the one of the head recognition system that worked on photographs taken in one session? Funnily enough its performance crashed when taking new photographs... it was recognizing the clothes.
> And even if you were ... careless ... enough to only use photos > sourced from the same places for all training, dry run, and demoing purposes, surely > you had some way to inspect the program to see why it was making the choices it made > (i.e. highlight on the photo those pixels which contributed to the friend/foe > decision), and would have seen right away that the thing was always focusing on the > leaf color or shiny surfaces facing west or whatever.
If you're capable of saying what a trained neural network actually bases its decisions on, there are some machine learning papers you should write that would be appreciated by the community.
OG.
|
|