Károly Zsolnai-Fehér covers a statistical algorithm that transfers writing style from one text to another without neural networks. It analyzes how a source text changes style, then applies that pattern to new writing. The method outperforms neural approaches and could help graphic designers automate style changes.
Dear fellow scholars, this is two minute papers with before we start, it is important to emphasize that this paper is not using neural networks. Not so long ago in 2015, the news took the world by storm. Researchers were able to create a novel neural network based technique for artistic style transfer which had quickly become a small subfield of its own within machine learning. The problem definition was the we provide an input image and a source photograph and the goal is to extract the artistic style of this photo and apply it to our image. The results were absolutely stunning, but at the same time it was difficult to control the outcome. Later this technique was generalized for higher resolution images and instead of waiting for hours, it now works in almost real time and is used in several commercial products. Wow. It is rare to see a new piece of technology introduced to the market so quickly. Really cool. However, this piece of work showcases a handcrafted algorithm that only works on a specialized case of inputs text based effects, but in this domain it smokes the competition. And here the style transfer happens not with any kind of neural network or other popular learning algorithm, but in terms of statistics. In this formulation we know about the source text as well, and because of that we know exactly the kind of effects that are applied to it. Kind of like a before and after image for some beauty product if you will. This opens up the possibility of analyzing its statistical properties and applying a similar effect to practically any kind of text input. The term statistically means that we are not interested in one isolated case, but we describe general rules, namely in what distance from the text, what is likely to happen to it. The resulting technique is remarkably robust and works on a variety of input output pairs and is head and shoulders beyond the competition, including the state of the art neural network based techniques that is indeed quite remarkable. I expect graphic designers to be all over this technique in the very near future. This is an excellent, really well written paper and the evaluation is also of high quality. If you wish to see how one can do this kind of magic by hand without resorting to neural networks, don't miss out on this one and make sure to have a look. There's also a possibility of having a small degree of artistic control over the outputs, and who knows, some variant of this could open up the possibility of a fully animated style transfer from one image. Wow. And before we go, we'd like to send a huge shout out to our fellow scholars who contributed translations to our episodes for a variety of languages. Please note that the names of the contributors are always available in the video. Description it is really great to see how the series is becoming more and more available for people around the globe. If you wish to contribute, click on the cogwheel icon in the lower right and the subtitle CC text. Thank you so much. Thanks for watching and for your generous support, and I'll see you next time.
