9 Cool Deep Learning Applications | Two Minute Papers #35

9 Cool Deep Learning Applications | Two Minute Papers #35

November 23, 2019 29 By Stanley Isaacs


Dear Fellow Scholars, this is Two Minute Papers
with Károly Zsolnai-Fehér. There are so many applications of deep learning,
I was really excited to put together a short, but really cool list of some of the more recent
results for you Fellow Scholars to enjoy. Machine learning provides us an incredible
set of tools. If you have a difficult problem at hand, you don’t need to hand craft an algorithm
for it. It finds out by itself what is important about the problem and tries to solve it on
its own. In many problem domains, they perform better than human experts. What’s more, some
of these algorithms find out things that could earn you a PhD with 10 years ago. Here goes the first stunning application:
toxicity detection for different chemical structures by means of deep learning. It is
so efficient that it could find toxic properties that previously required decades of work by
humans who are experts of their field. Next one. Mitosis detection from large images.
Mitosis means that cell nuclei are undergoing different transformations that are quite harmful,
and quite difficult to detect. The best techniques out there are using convolutional
neural networks and are outperforming professional radiologists at their own task. Unbelievable. Kaggle is a company that is dedicated to connecting
companies with large datasets and data scientists who write algorithms to extract insight from
all this data. If you take only a brief look, you see an incredibly large swath of applications
for learning algorithms. Almost all of these were believed to be only for humans, very
smart humans. And learning algorithms, again, emerge triumphant on many of these. For example, they had a great competition
where learning algorithms would read a website and find out whether paid content is disguised
there as real content. Next up on the list: hallucination or sequence
generation. It looks at different video games, tries to learn how they work, and generates
new footage out of thin air by using a recurrent neural network. Because of the imperfection of 3D scanning
procedures, many 3D scanned furnitures that are too noisy to be used as is. However, there
are techniques to look at these really noisy models and try to figure out how they should
look by learning the symmetries and other properties of real furnitures. These algorithms can also do an excellent
job at predicting how different fluids behave in time, and are therefore expected to be
super useful in physical simulation in the following years. On the list of highly sophisticated scientific
topics, there is this application that can find out what makes a good selfie and how
good your photos are. If you really want to know the truth. Here is another application where a computer
algorithm that we call deep q learning, plays pong, against itself, and eventually achieves
expertise. The machines are also grading student essays.
At first, one would think that this cannot possibly be a good idea. As it turns out,
their judgement is more consistent with the reference grades than any of the teachers
who were tested. This could be an awesome tool for saving a lot of time and assisting
the teachers to help their students learn. This kind of blows my mind. It would be great
to take a look at an actual dataset if it is public and the issued grades, so if any
of you Fellow Scholars have seen it somewhere, please let me know in the comments section! These results are only from the last few years,
and it’s really just scratching the surface. There are literally hundreds of more applications
we haven’t even talked about. We are living extremely exciting times indeed. I am eager
to see, and perhaps, be a small part of this progress. There are tons of reading and viewing materials
in the description box, check them out! Thanks for watching and for your generous
support, and I’ll see you next time!