Piano Genie Trained a Neural Net to Play 88-Key Piano with 8 Arcade Buttons

Want to sound great on a Piano using only your coding skills? Enter Piano Genie, the result of a research project from Google AI and DeepMind. You press any of eight buttons while a neural network makes sure the piano plays something cool — compensating in real time for what’s already been played.

Almost anyone new to playing music who sits down at a piano will produce a sound similar to that of a cat chasing a mouse through a tangle of kitchen pots. Who can blame them, given the sea of 88 inexplicable keys sitting before them? But they’ll quickly realize that playing keys in succession in one direction will produce sounds with consistently increasing or decreasing pitch. They’ll also learn that pressing keys for different lengths of times can improve the melody. But there’s still 88 of them and plenty more to learn, such as which keys will sound harmonious when played together.

Piano Genie training architectureWith Pinao Genie, gone are the daunting 88 keys, replaced with a 3D-printed box of eight arcade-style buttons which they made by following this Adafruit tutorial. A neural network maps those eight buttons to something meaningful on the 88-key piano keyboard. Being a neural network, the mapping isn’t a fixed one-to-one or even one-to-many. Instead, it’s trained to play something which should sound good taking into account what was play previously and won`t necessarily be the same each time.

To train it they use data from the approximately 1400 performances of the International Piano e-Competition. The result can be quite good as you can see and hear in the video below. The buttons feed into a computer but the computer plays the result on an actual piano.

For training, the neural network really consists of two networks. One is an encoder, in this case a recurrent neural network (RNN) which takes piano sequences and learns to output a vector. In the diagram, the vector is in the middle and has one element for each of the eight buttons. The second network is the decoder, also an RNN. It’s trained to turn that eight-element vector back into the same music which was fed into the encoder.

Once trained, only the decoder is used. The eight-button keyboard feeds into the vector, and the decoder outputs suitable notes. The fact that they’re RNNs means that rather than learning a fixed one-to-many mapping, the network takes into account what was previously played in order to come up with something which hopefully sounds pleasing. To give the user a little more creative control, they also trained it to realize when the user is playing a rising or falling melody and to output the same. See their paper for how the turned polyphonic sound into monophonic and back again.

If you prefer a different style of music you can train it on a MIDI collection of your own choosing using their open-sourced model. Or you can try it out as is right now through their web interface. I’ll admit, I started out just banging on it, producing the same noise I would get if I just hammered away randomly on a piano. Then I switched to thinking of making melodies and the result started sounding better. So some music background and practice still helps. For the video below, the researcher admits to having already played for a few hours.

This isn’t the first project we’ve covered by these Google researchers. Another was this music synthesizer again using neural networks but this time with a Raspberry Pi. And if our discussion of recurrent neural networks went a bit over your head, check out our overview of neural networks.

Continue reading “Piano Genie Trained a Neural Net to Play 88-Key Piano with 8 Arcade Buttons”

Up AlphaGoer Five

AlphaGo is the deep learning program that can beat humans at the game Go. You can read Google’s highly technical paper on it, but you’ll have to wade through some very academic language. [Aman Agarwal] has done us a favor. He took the original paper and dissected the important parts of in in plain English. If the title doesn’t make sense to you, you need to read more XKCD.

[Aman] says his treatment will be useful for anyone who doesn’t want to become an expert on neural networks but still wants to understand this important breakthrough. He also thinks people who don’t have English as a first language may find his analysis useful. By the way, the actual Go matches where AlphaGo beat [Sedol] were streamed and you can watch all the replays on YouTube (the first match appears below).

Continue reading “Up AlphaGoer Five”

Catastrophic Forgetting: Learning’s Effect on Machine Minds

What if every time you learned something new, you forgot a little of what you knew before? That sort of overwriting doesn’t happen in the human brain, but it does in artificial neural networks. It’s appropriately called catastrophic forgetting. So why are neural networks so successful despite this? How does this affect the future of things like self-driving cars? Just what limit does this put on what neural networks will be able to do, and what’s being done about it?

The way a neural network stores knowledge is by setting the values of weights (the lines in between the neurons in the diagram). That’s what those lines literally are, just numbers assigned to pairs of neurons. They’re analogous to the axons in our brain, the long tendrils that reach out from one neuron to the dendrites of another neuron, where they meet at microscopic gaps called synapses. The value of the weight between two artificial neurons is roughly like the number of axons between biological neurons in the brain.

To understand the problem, and the solutions below, you need to know a little more detail.

Continue reading “Catastrophic Forgetting: Learning’s Effect on Machine Minds”

Talking Neural Nets

Speech synthesis is nothing new, but it has gotten better lately. It is about to get even better thanks to DeepMind’s WaveNet project. The Alphabet (or is it Google?) project uses neural networks to analyze audio data and it learns to speak by example. Unlike other text-to-speech systems, WaveNet creates sound one sample at a time and affords surprisingly human-sounding results.

Before you rush to comment “Not a hack!” you should know we are seeing projects pop up on GitHub that use the technology. For example, there is a concrete implementation by [ibab]. [Tomlepaine] has an optimized version. In addition to learning English, they successfully trained it for Mandarin and even to generate music. If you don’t want to build a system out yourself, the original paper has audio files (about midway down) comparing traditional parametric and concatenative voices with the WaveNet voices.

Another interesting project is the reverse path — teaching WaveNet to convert speech to text. Before you get too excited, though, you might want to note this quote from the read me file:

“We’ve trained this model on a single Titan X GPU during 30 hours until 20 epochs and the model stopped at 13.4 ctc loss. If you don’t have a Titan X GPU, reduce batch_size in the train.py file from 16 to 4.”

Last time we checked, you could get a Titan X for a little less than $2,000.

There is a multi-part lecture series on reinforced learning (the foundation for DeepMind). If you wanted to tackle a project yourself, that might be a good starting point (the first part appears below).

Continue reading “Talking Neural Nets”

Ask Hackaday: Google Beat Go; Bellwether or Hype?

We wake up this morning to the news that Google’s deep-search neural network project called AlphaGo has beaten the second ranked world Go master (who happens to be a human being). This is the first of five matches between the two adversaries that will play out this week.

On one hand, this is a sign of maturing technology. It has been almost twenty years since Deep Blue beat Gary Kasparov, the reigning chess world champion at the time. Although there are still four games to play against Lee Sedol, it was recently reported that AlphaGo beat European Go champion Fan Hui in five games straight. Go is generally considered a more difficult game for machine minds to play than chess. This is because Go has a much larger pool of possible moves at any given time.

Does This Matter?

Okay, the news part of this event has been covered: machine beats man. Does it matter? Will this affect your life and how? We want to hear what you think in the comments below. But I’m going to keep going with some of my thoughts on the topic.

You're still better at Ms. Pacman [Source: DeepMind paper in Nature]
You’re still better at Ms. Pacman [Source: DeepMind paper in Nature]
Let’s look first at what AlphaGo did to win. At its core, the game of Go is won by figuring out where your opponent will likely make a low-percentage move and then capitalizing on that choice. Know Your Enemy has been a tenet of strategy for a few millennia now and it holds true in the digital age. In addition to the rules of the game, AlphaGo was fed a healthy diet of 30 million positions from expert games. This builds behavior recognition into the system. Not just what moves can be made, but what moves are most likely to be made.

DeepMind, the company behind AlphaGo which was acquired by Google in 2014, has published a paper in Nature about their approach. They were even nice enough to let us read without dealing with a paywall. The secret sauce is the learning process which at its core tries to mimic how living entities learn: observe repetitively while assigning values to outcomes. This is key as it leads past “intellect”, to “intelligence” (the “I” in AI that everyone seems to be waiting for). But this is a bastardized version of “intelligence”. AlphaGo is able to recognize and predict behavior, then make choices that lead to a desired outcome. This is more than intellect as it does value the purpose of an opponent’s decisions. But it falls short of intelligence as AlphaGo doesn’t consciously understand the purpose it has detected. In my mind this is exactly what we need. Truly successful machine learning will be able to make sense out of sometimes irrational input.

The paper from Nature doesn’t go into details about Go, but it explains the approach of the learning system applied to Atari 2600. The algorithm was given 210×160 color video at 60Hz as an input and then told it could use a joystick with one button. From there it taught itself to play 49 games. It was not told the purpose or the rules of the games, but it was given examples of scores from human performance and rewarded for its own quality performances. The chart above shows that it learned to play 29 of them at or above human skill levels.

Continue reading “Ask Hackaday: Google Beat Go; Bellwether or Hype?”