The idea is that you can predict a set from any conditioning by: 1. Encoding the conditioning. 2. Encoding a randomly initialised set of some maximum size. 3. Running gradient descent on the set to match its encoding to the encoding of the conditioning.
-
-
Prikaži ovu nitHvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Not sure I understood but I expected something different from the notion of set. Nothing related to images.
-
I guess the authors are vision-oriented, but the same principle can be used to predict any set, e.g. 3D point cloud, a set of users using a service at a particular time, or a set of products of a chemical reaction.
- Još 2 druga odgovora
Novi razgovor -
-
-
What's the idea?
-
The idea is that you can predict a set from any conditioning by: 1. Encoding the conditioning. 2. Encoding a randomly initialised set of some maximum size. 3. Running gradient descent on the set to match its encoding to the encoding of the conditioning.
Kraj razgovora
Novi razgovor -
-
-
This paper reminds me of the GLO paper https://arxiv.org/abs/1707.05776 where you can do inference in GANs (i.e. map image to the latent z) by running gradient descent on z to match the image. Applying this idea to infer sets is clever.
@Cyanogenoid perhaps worth citing GLO? -
Good find, thanks
@hyunjik11!
Kraj razgovora
Novi razgovor -
-
-
I'm super happy to hear that you liked it this much! This made my day!
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Really interesting paper although for me I would be curious to see how it scales on harder problems (for example higher dimensional sets). It feels for me like re-learning the representation through gradient descent might not be computationally feasible for some complex problems.
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
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.