3 ms·

Author mentions style transfer as a product of NN research. Yet this paper [1] from 2014 works extremely well, no neural networks. Here's one from 2008 [2] tha

Author mentions style transfer as a product of NN research. Yet this paper [1] from 2014 works extremely well, no neural networks. Here's one from 2008 [2] that worked extremely well too, better than most early NN techniques. Although, deep learning is far from the intuitive approach that existed 10 years ago. It's clearer now how to reason about models, layers, activation functions etc. As much as there's no mathematical foundation, I do not believe it helps that much in the case of SVMs or whatever else. There still has to be experimentation, approximation and proper testing. [1]: https://dspace.mit.edu/handle/1721.1/100018 https://dspace.mit.edu/handle/1721.1/100018 [2]: http://link.springer.com/chapter/10.1007%2F978-3-540-88690-7_3?LI=true http://link.springer.com/chapter/10.1007%2F978-3-540-88690-7...

jph00 9y ago
To claim that those techniques are even in the same ballpark as deep learning based style transfer is just silly. They are extremely limited in their application, as you can clearly see by looking at the examples. And the entire approaches are overfit to the particular applications that they show, so they can't be readily extended. These are actually excellent examples of how deep learning has totally changed our expectations of what can be done with computer vision by using deep learning.
- projectorlochsa 9y ago
  Deep learning based style transfer that got viral a year or two ago was worse than this 2008 result, especially worse for that hone application to portraits. The hype was nowhere near just because there's no deepness. I admit that https://arxiv.org/abs/1705.01088 https://arxiv.org/abs/1705.01088 this beats everything and is extremely powerful and simple but the hype for deep things seems a bit too strong.
j2kun 9y ago
It's important to note that we know the limitations of SVMs because of the theory. That's usually what a theory provides. There are also a few methods, like boosting, that are purely the result of theory.
j2kun 9y ago
It's important to note that we know the limitations of SVMs because of the theory. That's usually what a theory provides. There are also a few methods, like boosting, that are purely the result of theory.

jph00 9y ago

To claim that those techniques are even in the same ballpark as deep learning based style transfer is just silly. They are extremely limited in their application, as you can clearly see by looking at the examples. And the entire approaches are overfit to the particular applications that they show, so they can't be readily extended. These are actually excellent examples of how deep learning has totally changed our expectations of what can be done with computer vision by using deep learning.

projectorlochsa 9y ago
Deep learning based style transfer that got viral a year or two ago was worse than this 2008 result, especially worse for that hone application to portraits. The hype was nowhere near just because there's no deepness. I admit that https://arxiv.org/abs/1705.01088 https://arxiv.org/abs/1705.01088 this beats everything and is extremely powerful and simple but the hype for deep things seems a bit too strong.

j2kun 9y ago

It's important to note that we know the limitations of SVMs because of the theory. That's usually what a theory provides. There are also a few methods, like boosting, that are purely the result of theory.

j2kun 9y ago