Fortunately, we are kept busy working on client projects most of the time. But during the gaps in between, rather than twiddle our collective thumbs we try to put them to good use on internal developments and initiatives. Free of a client brief, budget, and deadline, we really do have a blank canvas. So, what should we do with it?
Here in the Wireless and Digital Services (WDS) division of our Boston office, we want to explore new areas of machine learning and audio, inspired by our colleagues’ work on The Aficionado, smarter recycling and Vincent.
We’re not out to create a marketable product or even to solve a particular problem. We simply want to push ourselves as engineers to develop something unique and engaging. We will be recording our progress in a series of blogs, starting right here.
Our goals are to:
- Create something a ‘wow factor’ that people can engage with
- Explore and develop our machine learning and audio capabilities
And we’re off!
Our first step was to generate a long-list of ideas. We did some independent research before getting together to discuss our ideas and draw them up on the whiteboard walls.
We took an ‘anything goes’ approach – everyone in the room could suggest ideas focused on machine learning, audio, or a combination of the two. Our attempts at drawing definitely demonstrated that none of us are talented artists. (Perhaps we should enlist the help of Vincent next time round!)
Ideas in the longlist included:
- Automatic music accompaniment – use deep learning to automatically generate accompaniment in real-time during a musical performance
- Smart echo cancelation – replace algorithm-based echo cancelation methods with neural networks
- Miniature self-driving car – turn a remote-control car into a miniature self-driving vehicle using machine learning and machine vision techniques
And the winners are…
Using a dot-voting method, three winning ideas were selected:
- FoosBot (Fil Dourado) – use cutting-edge Deep Reinforcement Learning to train an agent to play foosball against a human competitor. The game would take place in a virtual reality environment while players use physical handles to provide input to the game.
- Sunk melodies (Amy Stephens) – explore the feasibility of hiding recoverable information in audio signals. Examine the use of audio side-channels for authentication, key exchange and audio QR codes.
- Dereverberation with neural networks (Jay Biernat) – train a series of neural network to do deconvolution of room reverberation with minimal artefacts. Experiment with providing approximate room impulse responses as additional input to improve results.
The Aficionado is a machine learning system for the infinitely complex world of music
The great thing about all of these ideas is that they require novel solutions to be combined with good engineering practice to deliver a compelling result. Our focus is on taming new and exciting technologies to realise practicable, real-world solutions.
Jay, Fil, and Amy are now researching and developing their ideas in preparation to present them, elevator pitch style, to the WDS senior team in February.
Tune in next time for more details on their ideas and to find out how they get on.