Capsule Networks- the New Deep Learning Network | Hacker News Reader

Capsule Networks- the New Deep Learning Network

towardsdatascience.com

38 pointsjoeyespo8 comments

Aren't Capsule Networks one of this hyped ideas - everyone was talking about them (because their rationale sounded interesting, and well, Hinton) were never seen again?

It is very different to good ideas, like 3x3 conv, dropout, batch norm, residual connections - which once introduced, become standard building block of networks.

They are not hyped. Capsule nets are in early phase of research.

The bleeding edge deep learning research that tries to push the science forward is trying to develop new ideas. Hinton et. al developed the current deep learning revolution gradually over decades. Next non-incremental revolutions in the field may take similar time to mature.

Oh for sure! Not everybody needs to be trying to max out the standard benchmarks with every paper. But have there been any new developments or improvements with capsule nets since the original paper? This blog post looks like it only references that first one, which is coming up on two years old, right? It'd be great to read about any followup.

Graph Capsule Convolutional Networks [0] claim to have beaten the state of the art in graph neural networks. However, I can’t find information on training time, which I think is significantly more expensive for capsule networks.

[0] https://arxiv.org/abs/1805.08090

There are 4 problems that need to be solved: translation, rotation, reflection and dilation. Capsule networks are potentially one way to do that.

> but because CNN’s only look for features in images, and don’t pay attention to their pose, it’s hard for them to notice a difference between that face and a real face.

Then this makes me wonder: how are GANs able to create pictures of people where eyes and noses and so on are correctly placed with respect to eachother?

E.g. as in: https://www.theverge.com/2018/12/17/18144356/ai-image-genera...

GANs are a battle between the discriminator and the generator. If the G places eyes in the wrong place, the D and G eventually get error signals that cause them to improve. Using pooling layers, CNNs build features that look across the entire image. So it can learn that eyes need to go across each other and above the nose. It is just that when used for classification, CNNs may not need to learn this because it is never trained to discriminate real and fake faces.

I am a fan of giving new nicknames to things when I do any noticeable, even collateral, in extremely early stage, improving.

Financially wise, must be a good thing to do, too, but it is not the main point. The main point is Not that I sound smarter when I use the new nickname(therefore this is not exactly sarcasm), but that I can have a more-at-hand/shorter way to identify the new, subtly improved version, and It saves a lot of words, in the "spoken". colloquial, communication.

Now for the matters of Science improvement itself, I mean, like for Humanity and stuff, I am not clear if this is good, maybe. Science and business, in a very subtle, almost invisible way, mixed. Always have been like this, and it probably will continue like this, for a while.

Reply on news.ycombinator.com