you are viewing a single comment's thread.

view the rest of the comments →

[–]magnora7[S] 2 insightful - 1 funny2 insightful - 0 funny3 insightful - 1 funny -  (3 children)

Yeah, exactly. It's able to resolve the image details on the feature level, instead of just the whole-image level. Then it's able to combine those features (nose, eye, chin, etc) to make new faces, but also keep everything consistent on a whole-image level. So it's able to make faces that never existed and everything in between.

Basically we're seeing a memory dump of the basis vectors of the neural network. It has learned these faces, and then any new faces will be judged as being (for example) 0.5 of this face, 0.3 of that face, and 0.2 of that face. (And 0.2 of this nose, and 0.4 of that nose... and so on). And this is how the neural net "remembers" stuff and can recall stuff. So we're seeing the 10 faces or so it resolved as "example faces" to use as basis vectors, with the transitions between them since it's a continuous function.

I hope I didn't over-explain, but neural nets and genetic algorithms working together is something I've spent years working on. It's so fascinating to me, I almost released an indie unity game based around genetic algorithm neural networks but I never finished it.

[–]d3rr 2 insightful - 1 funny2 insightful - 0 funny3 insightful - 1 funny -  (2 children)

Okay, I'm on track then. Nope, great explanation. I read a book about neural nets and genetic algorithms but I haven't kept up with anything newer, or implemented anything in code. It is pretty fascinating though, with the engineers not being able to explain how a particular result was arrived at.

That's awesome that you worked on a game with this stuff. I remember when Spore was coming out everyone thought it might really be something special. I never played it actually, but people didn't seem too excited about the end result. I was messing around with a 2D tile based RPG myself before this whole thing kicked off... didn't get too far but got some cool stuff in place and demo-able.

[–]magnora7[S] 1 insightful - 1 funny1 insightful - 0 funny2 insightful - 1 funny -  (1 child)

That sounds like a cool game idea. My game was basically a "set the parameters and wait" kind of game, like seeing what type of neural net you could evolve that would move a 3d character with 10 joints or whatever, each output mapping to a joint. Then they'd be rated on their fitness to complete some task, like jumping the highest or walking the farthest.

[–]d3rr 2 insightful - 1 funny2 insightful - 0 funny3 insightful - 1 funny -  (0 children)

Oh that sounds awesome, even if it turned into more of a simulation than a game.