DURHAM, N.C.—Deep Dream was a nightmare. Google's experiment in applying neural networks to images has been used within Google to greatly improve search and recognition. But for most of us, it's also associated with a horrifying image filter that mixes together everything a neural network thinks it might possibly recognize in an image—filling a picture of kittens with eyeballs and spiders, for instance.
But neural networks, computer programs that can learn from examples, are here to stay. They're a great way for computers to recognize things they aren't strictly programmed to understand. Self-driving cars, for example, can use neural networks for video analysis and scene recognition, understanding signs, obstacles, and things on the road. Here at Moogfest, Google showed off another project from its Google Brain division: Magenta, which applies the neural network idea to composing music.
"We're exploring how we might imbue machines with creativity using machine learning," Google researcher Adam Roberts said. "Can it create music as well?"
Recommended by Our Editors
Magenta is more complex than Deep Dream, because music has an element of time that isn't necessary for images, Roberts said. So now the neural network needs to remember, linking together the notes it previously played with the ongoing melody.
"Over time, you're remembering what you saw before, so there's this little history the neural network is keeping and can use to produce music," Roberts said. Magenta will be open sourced "very soon," on Github.
"It's basically just seeing what I play, trying to find what's the most probably note to come after that based on the music we played for it, and again, this is a very early start," Roberts said.
Watch the video above to see Magenta's first public debut, using the Bob Moog Google Doodle to create its own original composition. Frankly, it's pretty rough stuff. Neural networks start out not knowing very much, after all. Their power is in their ability to learn.