Deep Learning fags BTFO

What makes you think your brain works differently?

>>8656904
Winter IA.

>>8656904
dude the comeputer is like a replica of the brain man

>>8656923

This. 99.99% of things I do everyday I know how to do by heart. I haven't learned them, I have memorized a huge fucking shit load of puzzle pieces that I can put together to solve a huge amount of situations.

also 99.99% of the time the pieces go together in almost the same way every time.

>>8656938
That's utterly and completely missing the point. People want these kinds of systems to be able to GENERALIZE i.e. be able to tell that a cat viewed from a different angle or some shit from the cats it was trained on is still a cat. These guys do a bunch of experiments to show that what they more likely do is memorize and do shallow comparisons.

Not surprising really, this kind of supervised CNN-type crap has always rubbed me the wrong way. Hopefully this gets people thinking more about GANs, VAEs, and other methods of learning that are more subtle than shoving a million labeled images through a network.

Glad this was shown before they started using these to diagnose cancer or some shit.

>>8656977
blease describe GANS and VAES
Also, couldnt the shallow comparisons be refined later?

>>8656904
hardly. The authors still admit that "large neural networks generalize well in practice."

>>8656977
GANS are a type of deep learning though. CNNs are on their way out already because adversarial examples are a problem.

Deep learning a shit. Hand-crafted features are where it's at.
This once again proves that only human is capable of making breakthroughs.

>>8657027
>GANS are a type of deep learning though.
Yeah but the way they're designed helps deter this kind of thing. If the discriminator is just memorizing then the generator will pick up on it and create data specifically to exploit that weakness, thereby getting rid of it. Theoretically anyway.

>>8657042
GAN main purpose is not to train the discriminator but the generator. The discriminator is deliberately made weak so that the discriminator can learn experience from interacting with it.

File: confused.jpg (8KB, 300x256px) Image search: [Google]

8KB, 300x256px

>>8656904
>Deep neural networks work extremely well for certain tasks
>We keep finding new uses for them
>Deep Learning fags BTFO

>>8656904
>including a tl;dr in a research paper
Is this a joke?

>>8656904
Yes, we know that there is a lot of theoretical bleakness in ML. That's why it's more important to focus on the practical successes while we can still leverage improved computing power.

File: 1479266334823.jpg (10KB, 157x157px) Image search: [Google]

10KB, 157x157px

You can't fool me. I know the practical successes of CNNs are too high and are going to replace countless millions of jobs and be worth billions. You're trying to minimise interest to keep those billions to yourself but I see right through your tricks

>>8657797

Not to mention that you can build GANs with CNNs... this entire discussion is like, 'LOL WE /truck/ NOW! TEAR OUT THAT ENGINE!!!'

File: alphago-vs-lee-sedol-2_w_600.jpg (43KB, 600x344px) Image search: [Google]

43KB, 600x344px

>>8656904
Hi. You are wrong.