Have you actually published the ANML work yet? Currently working on something similar, but not clear about some of your details.
-
-
-
As you say, it is not evolution because it is a different optimization algorithm. It is analogous to evolution in some ways, but different in others. Yes, I believe in outer loop can be truly open-ended. See our POET work, which moves in that direction.https://eng.uber.com/poet-open-ended-deep-learning/ …
Kraj razgovora
Novi razgovor -
-
-
Excellent talk today
@jeffclune ! Felt like I am watching a big moment in history unfold. -
Thank you!
Kraj razgovora
Novi razgovor -
-
-
Looking forward the recording!
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Always excited to welcome you back! :)
#reworkdlHvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
hmm I find this part a bit weird: https://slideslive.com/38921888/biological-and-artificial-reinforcement-learning-3?t=3216 … after training in inner loop on class T, you eval on other classes T'. You are meta-training an algorithm that knows about the distribution of classes, rather than directly optimizing to avoid interference among classes?
-
ah on the OML paper, they define the omniglot meta-training train set as a sequence of images,5 from each of a sequence of 200 classes. ah,I just found the trick they used, which is what you meant on the talk. It's like truncated backprop through time, but on a sequence of tasks
- Još 1 odgovor
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.