News for hackers. Stuff that matters.
Re:World Models by wei_jok on Tuesday March 27, 2018 @08:52PM (#16694209)
Link to the neural net generated Doom level demo:

https://worldmodels.github.io/doomrnn/

Original article:

https://worldmodels.github.io/

PDF version of the paper:

https://arxiv.org/abs/1803.10122

Re:World Models by mh-cx on Wednesday March 28, 2018 @04:33AM (#16696307)
Hmm, all I can see is a brownish blurry ... uhm ... something (Chrome, Desktop). Could someone explain what we are expected to see here?
Re:World Models by gnode on Wednesday March 28, 2018 @09:08AM (#16697367)
As I understand it, it's supposed to be a blurry dream-like vision of Doom (or similar game) hallucinated by a neural network.
Re:World Models by kroltan on Wednesday March 28, 2018 @08:07AM (#16697041)
Same here but with Firefox 58, Desktop, Intel graphics.

Additionally, the brownish blur has some additional variation at the top of the screen.

Re:World Models by autogeek on Wednesday March 28, 2018 @08:22AM (#16697117)
BEWARE: This thing eats memory like crazy on firefox.

I'm clocking more than 2GB memory usage for this site after 30 secs on firefox.

EDIT: Seems to be firefox specific. Lots of CPU use on Chrome but memory consumption seems to be under control.

Re:World Models by eli_gottlieb on Tuesday March 27, 2018 @09:47PM (#16694513)
So it's basically a recurrent neural-net version of model-predictive control, applied to reinforcement learning problems?

Nice.

Re:World Models by chirau on Wednesday March 28, 2018 @09:47AM (#16697562)
Lol @ 'basically'
Re:World Models by eli_gottlieb on Wednesday March 28, 2018 @02:48PM (#16700581)
Working in the cognitive sciences is very weird, because you can almost always point to an idea in neuroscience and cognitive psychology, to something about the real brain, that makes a new machine learning paper look downright trivial. Simultaneously, actually implementing even established ideas about cognition from the meat-side to the software-side tends to take a lot of work.

Admittedly, IMHO, this is often because machine learners are constantly trying to transform any idea from neuroscience or cognitive science into an idea about supervised loss minimization in a neural network, runnable on a graphics card, rather than having to come up with ways of efficiently computing the stuff that we're confident the brain actually computes.

Re:World Models by johnhenry on Tuesday March 27, 2018 @10:21PM (#16694701)