10 More Cool Deep Learning Applications | Two Minute Papers #52

10 More Cool Deep Learning Applications | Two Minute Papers #52

November 26, 2019 27 By Stanley Isaacs


Dear Fellow Scholars, this is Two Minute Papers
with Károly Zsolnai-Fehér. To all of you Fellow Scholars out there who
are yearning for some more deep learning action like I do, here goes the second package. Buckle
up, amazing applications await you. As always, links to every one of these works
are available in the description box. This convolutional neural network can learn
how to colorize by looking at the same images both in colo r and black and white. The first image is the black and white input,
the second is how the algorithm colorized it, and the third is how the image originally
looked like in color. Insanity. Recurrent neural networks are able to learn
and produce sequences of data, and they are getting better and better at music generation.
Nowadays, people are experimenting with human aided music generation with pretty amaz ing
results. Sony has also been working on such a solution
with spectacular results. One can also run a network on a large database
of leaked human passwords and try to crack new accounts building on that knowledge. Deep neural networks take a substantial amount
of time to train, and the final contents of each of the neurons have to be stored, which
takes a lot of space. New techniques are being explored to compress the information content
of these networks. There is an other application where endangered
whale species are recognized by convolutional neural networks. Some of them have a worldwide
population of less than 500, and this is where machine learning steps in to try to save them.
Awesome! YouTube has a huge database full of information
on what kind of video thumbnails are the ones that people end up clicking on. They use deep
learning to automatically find and suggest the most appealing images for your videos. There is also this crazy application where
a network was trained on a huge dataset with images of celebrities. A low quality
image is given, where the algorithm creates a higher resolution version building on this
knowledge. The leftmost images are the true high resolution images, the second one is
the grainy, low resolution input, and the third is the neural network’s attempt to reconstruct
the original. This application takes your handwriting of
a number, and visualizes how a convolutional neural network understands and classifies
it. Apparently, George RR Martin is late with
writing the next book of Game of Thrones, but luckily, we have recurrent neural networks
that can generate text in his style. An infinite amount, so beware George, winter is coming.
I mean the machines are coming. It is truly amazing what these techniques
are capable of. And as machine learning is a remarkably fast moving field, new applications
pop up pretty much every day. I am quite enthused to do at one more batch of these! Of course,
provided that you liked this one. Let me know. Thanks for watching and for your generous
support, and I’ll see you next time!