Комментарии:
Uh. Umm. Uhh. Ummm
ОтветитьThank you Jeff Dean.
ОтветитьI would be curious to see if, going along with this 'learning to learn' task, that we could use gans to generate similar, but different problems. RL could be used to keep the difference between two problems in check and start to recognize when it can challenge itself in taking on new tasks more.
For example, I'm thinking about like going from image classification of like 100 x 100, then a GAN tells the RL NAS to go to 101 x 101. Next it generates a problem with two frames (a video), then it starts learning speech and sound source detection. Building on itself to increase data efficiency.
He may be like the most intelligent person you’ve never heard of, and I salute the incredible things he has done at google...but he can’t present for shi.....
Ответитьawesome presentation and even better questions at the end
ОтветитьLearning to learn for sure
Ответитьgood
Ответить