If you click here, you’ll hear an audio track that sounds like a child playing the piano. It is, however, a machine playing the piano. Who does the machine belong to? Google of course, Google Brain to be precise. This is the first public result of Project Magenta, a research effort by Google to advance machine learning in music and art. It’s an amateurish start, but hey, that’s exactly how AI is supposed to start.
Like humans, AI learn over time, and theoretically, are supposed to master a subject eventually. Google’s AlphaGo algorithm, which beat Go champion, Lee Sedol recently, did so by training itself over countless sessions of Go. It also learned from other Go players around the world.
Google has been stepping up its efforts in the machine learning space for a while now. The company recently, even announced a ASIC specially designed for machine learning implementation. The company’s Tensor Processing Unit (TPUs) are specially designed for machine learning implementation. Google claims the performance per watt of these TPUs, is “an order of magnitude higher” than GPUs. According to a Google spokesperson, “TPU boards are the size of a disk drive, and are designed primarily for data center applications.”
That said, beating Go and regular speech recognition efforts may be simpler for machine learning, than music and art. To do this, an AI will have to learn what makes music worth listening to, or art worth perusing. Google is looking to get help from coders on GitHub as well, by making Project Magenta available on the code repository.