Like a good chunk of the community, I’ve spent some of my free time during the past few weeks playing around with char-rnn. There’s definitely something a little awe inspiring here. I won’t waist your time talking about how it works, Karpathy has already done an amazing job there. I wouldn’t just read through his blog post though, I’d urge you to take the time to read all the papers and other articles he’s linked to. I promise, you won’t regret it.
Some have said that it’s just regurgitating statistical information about the corpus it’s read through. While this is true I guess, it’s an attempt at removing the awe and wonder from it. It’s incredibly fun to see it work, it’s even interesting to those who don’t have a passion for machine learning. Friends were amazed to see it pump out text that looked a whole lot like Obama’s speeches, or Shakespeare, or music.
And on that note, this is a good segue to one of the most interesting uses for char-rnn I’ve tried, to analyze abc files and create for us beautiful beautiful music. Below you’ll find some examples of it working. It’s specifically interesting to note here that not only did it compose the song, it also titled it, which is pretty sweet.