Károly Zsolnai-Fehér explains how recurrent neural networks handle sequences of data, from image captions to machine translation. RNNs can generate letter-by-letter text that mimics Shakespeare novels, music, and Linux code, learning structure through training until the output looks convincing enough to fool readers.
Dear fellow scholars, this is two minute. Papers with artificial neural networks are very useful tools that are able to learn and recognize objects on images, or learn the style of Van Gogh and paint new pictures in his style. Today we're going to talk about recurrent neural networks. So what does the recurrent part mean? With an artificial neural network, we usually have a one to one relation between the input and the output. This means that one image comes in and one classification result comes out, whether the image depicts a human face or a train. With recurrent neural networks, we can have a one to many relation between the input and the output. The input would still be an image, but the output would not be a word, but a sequence of words, a sentence that describes what we see on the image. For a many to one relation, a good example is sentiment analysis. This means that a sequence of inputs from, for instance, a sentence is classified as either negative or positive. This is very useful for processing movie reviews where we'd like to know whether the user liked or hated the movie without reading pages and pages of discussion. And finally, recurrent neural networks can also deal with many to many relations, translating an input sequence into an output sequence. Examples of this can be machine translations that take an input sentence and translate it to an output sentence in a different language. For another example of a many to many relation, let's see what the algorithm learned after reading Tolstoy's War and Peace novel by asking it to write exactly in that style. It should be noted that generating a new novel happens letter by letter, so the algorithm is not allowed to memorize words. Let's look at the results at different stages of the training process. The initial results are, well, gibberish, but the algorithm seems to recognize immediately that words are basically a big bug bunch of letters that are separated by spaces. If we wait a bit more, we see that it starts to get a very rudimentary understanding of structures. For instance, a quotation mark that you have opened must be closed, and a sentence can be closed by a period and it is followed by an uppercase letter. Later it starts to learn shorter and more common words such as fall debt D. For me, if we wait for longer, we see that it already gets a grasp of longer words, and smaller parts of sentences actually start to make sense. Here's a piece of Shakespeare that was written by the algorithm. After reading all of his works, you see names that make sense and you really have to check the text thoroughly to conclude that it's indeed not the real deal. It can also try to write math papers. I had to look for quite a bit until I realized that some something is fishy here. It is not unreasonable to think that it can very easily deceive a non expert reader. Can you believe this? This is insanity. It is also capable of learning the source code of the Linux operating system and generate new code that looks quite sensible. It can also try to continue the song Let it Go from the famous Disney movie Froz. So recurrent neural networks are amazing tools that open up completely new horizons for solving problems where either the inputs or the outputs are not one thing, but a sequence of things. And now, signing off with a piece of recurrent neural network wisdom. Well, your wit is in the care of side that bear this in mind wherever you go. Thanks for watching and I'll see you next time.
