Earlier this week, AWS launched DeepComposer, a set of web-based AI learning tools for making music, and a $ 99 MIDI keyboard for inputting tunes. However, this start caused a bit of confusion. That's why we teamed up with Mike Miller, director of AWS 'AI Devices group, to discuss where DeepComposer fits into the company's AI device portfolio, which includes the DeepLens camera and the DeepRacer AI car, both are meant to teach developers specific AI concepts as well.
The first thing to remember is that DeepComposer is a learning tool. It's not meant for musicians – it's meant for engineers who want to learn about generative AI. However, AWS did not help, calling it "the world's first machine learning keyboard for developers". The keyboard itself is just a simple standard MIDI keyboard. There is no intelligence in it. All AI work takes place in the cloud.
"The goal is to teach generative AI as one of the most interesting trends in machine learning in the last 10 years," said Miller. "We've specifically told GANs, generative adversarial networks, that two networks are being trained together. The reason that we think is interesting for developers is that it's very complicated, and many things that developers learn about training machine learning models get confused when you train two together. "
With DeepComposer, the developer is pioneering a process of learning the basics. Use the keyboard to enter a basic melody. If you do not have them, you can also use an on-screen keyboard to get started, or use a few standard melodies (think of Ode to Joy). From a practical point of view, the system then creates a background track for this tune, based on a style of music you choose. However, for simplicity, the system ignores some values from the keyboard, including velocity (just in case you need more evidence that this is not a keyboard for musicians). More importantly, developers can then examine the system-generated models and even export them to a Jupyter notebook.
For DeepComposer, MIDI Data is Just Another Data Source Let developers know about GANs and SageMaker, the AWS machine learning platform that DeepComposer supports behind the scenes.
"The benefit of using MIDI files and introducing them to MIDI is that the data that goes into training is presented in a format that is, for example, the same representation of data in a picture," explained Miller. "And so it's actually very accurate and analogous, if a developer looks at this SageMaker notebook and understands the data formatting and how we share the data, it's applicable to other domains as well."
For this reason, the tools also provide all the raw data, including loss functions, analysis, and the results of the various models to produce an acceptable result. This is of course a tool for creating music. Also provide some data about the music such as pitch and empty bars.
"We believe that developers working on the SageMaker models will find that I can apply that data to other domains and take over it. Do it for yourself and see what I can generate," said Miller ,
Having heard the results so far, I can safely say that DeepComposer will soon stop producing hits. It seems pretty good to make a drum track, but the bass lines are a bit irregular. Still, it's a cool demo of this machine-learning technique, though I suspect that success will be slightly less than DeepRacer, a concept that is a little easier to understand for most, as the majority of developers are involved with it think they need to be able to play an instrument to use and keep going.
Additional coverage by Ron Miller.