Deep learning is doing wild things: finding locations from photos, upscaling tiny images, predicting what you'll type next, beating humans at Go. The field is moving so fast that new applications pop up every week.
Dear fellow scholars, this is two minute papers with this is the third episode in our series of deep learning applications. I have mixed in some recurrent neural networks for your and honestly my own enjoyment. I think this series of applications shows what an amazingly versatile tool we have been blessed with with deep learning. And I know you fellow scholars have been quite excited for this one. Let's get started. This piece of work accomplishes geolocation for photographs. This means that we toss in a photograph and it tells us exactly where it was made. Super resolution is a hot topic where we show a coarse, heavily pixelated image to a system and it tries to guess what it depicts and increase the resolution of it. If we have a tool that accomplishes this, we can zoom into images way more than the number of megapixels of our camera would allow. It is really cool to see that deep learning has also made an appearance in this subfield. This handy little tool visualizes the learning process in a neural network with the classical forward and backward propagation steps. This recurrent neural network continues our sentences in a way that kind of makes sense. Well, kind of Human in the loop techniques seek to create a bi directional connection between humans and machine learning techniques such as so they can both learn from each other. I think it definitely is an interesting direction at first. DeepMind's AlphaGo also learned the basics of go from amateurs and then took off like a hermit to learn on its own and came back with gun splazing. We usually have at least one remarkably rigorous and scientific application of deep learning in every collection episode. This time I'd like to show you this marvelous little program that suggests emojis for your images. It does so well that nowadays even computer algorithms are more hip than I am. This application is akin to the previous one we have seen about super resolution. Here we see beautiful high resolution images of digits created from these tiny, extremely pixelated inputs. Netflix is an online video streaming service. The Netflix Prize was a competition where participants wrote prize to estimate how a user would enjoy a given set of movies based on this user's previous preferences. The competition was won by an ensemble algorithm which is essentially a mixture of many existing techniques and by many I mean 107. It is not a surprise that some contemptuously use the term abomination instead of ensemble because of their egregious complexity. In this blog post a simple neural network implementation is described that achieves quite decent results and the core of the solution fits in no more than 20 lines of code. The code has been written using Keras, which also happens to be one of my favorite deep learning libraries. Wholeheartedly recommended for everyone who likes to code. And a big shout out to Francois, the developer of the mentioned library. Convolutional neural networks also have started curating works of art by assigning a score to how aesthetic they are. Oh, sorry, Leonardo. Earlier we talked about adversarial techniques that add a very specific type of noise to images to completely destroy the accuracy of previously existing image classification programs. The arms race has officially started and new techniques are popping up to prevent this behavior. If you find some novel applications of deep learning, just send the link my way in the comments section. Thanks for watching and for your generous support, and I'll see you next time.
