Super fun collaboration with @JohnFJMellor, @silverbottlep, @yaroslav_ganin, @ibabusch, @tejasdkulkarni, @danrsm, Andy Ballard, @theophaneweber and @OriolVinyalsML.
-
-
Prikaži ovu nit
-
When sufficiently constrained, agents learn to paint surprisingly abstract images. Some of the paintings remind me of cubist portraits. (Remember: no imitation or supervision). Can you spot any familiar faces? See https://learning-to-paint.github.io for loads more emergent drawing styles.pic.twitter.com/dDW5RnyAjh
Prikaži ovu nit -
When given more time with the canvas, agents produce images that look more natural. Of course, they're still constrained to draw with a brush, so their samples are not photo-real. But that's not the point.pic.twitter.com/O8wO8Y1sLQ
Prikaži ovu nit -
The point is that these agents' representations can actually be instantiated in physical reality. This is not necessarily the case with purely neural autoencoder representations. Read more about the video below here: https://deepmind.com/blog/article/learning-to-generate-images …pic.twitter.com/gPdp7QrrqC
Prikaži ovu nit -
This means that, in certain scenarios and under certain circumstances, the representations that these agents produce can be considered to be 'semantic'.pic.twitter.com/78Qr1wlWHr
Prikaži ovu nit -
The combination of: 1. learned generative agents, 2. physically plausible environments, and 3. learned adversarial reward functions, could be useful for program synthesis, inverse graphics, chemical synthesis, music generation and so much more.
Prikaži ovu nit -
And when it comes to straight-up creativity, I'd be very curious to hear what artists and art researchers of the likes of
@JoanneHastieArt,@dribnet,@patricktresset,@_ScottEaton_,@AaronHertzmann,@elluba make of this! Collaborations, anyone? ;)pic.twitter.com/49GbUmN1fl
Prikaži ovu nit -
@yaroslav_ganin recently open-sourced the RL environments we used for this work, check it out:http://github.com/deepmind/spiralPrikaži ovu nit
Kraj razgovora
Novi razgovor -
-
-
I was trying to progress with a similar idea for MRI acquisitions for a while, unfortunately, GAN with even with w-gp loss, never worked for me to any satisfactory standard. Your work motivates me to relook again. Thanks for sharing.
-
@arkitus, reading the paper more carefully, I will have an "out of memory" error as I don't have access to impala ;) Good work anyway!!
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.