Artificial Intelligence (AI) music composition research began decades ago, and since then, several organizations have been actively developing AI technology that can compose music without human intervention. Some startups, along with some well-established companies, have even used AI to create music that is commercially successful.

Sony’s AI system, Flow Machines, composed the song ‘Daddy’s Car’, which is a Beatles-style piece. They trained a machine learning algorithm on Beatles songs, which allowed the technology to compose a song based on what it had learned.

One of the most popular AI music composition startups is Aiva. This one-year-old company was founded in Luxembourg and London by Pierre Barreau, Denis Shtefan, Arnaud Decker, and Vincent Barreau. Its AI team is currently working on composing classical music, which is especially hard to do as it requires innate human emotion to compose the soulfulness for which classical music is known. Though the soundtracks are composed by AI, the orchestration, arrangement, and production of the music requires human intervention. (You can listen to one of their compositions here.)

Aiva’s musical tracks are commercially available for different industries interested in incorporating music in their products, ranging from film production companies to game studios. Aiva will create tracks upon request, providing every company with a customized, unique soundtrack for its product.

Deep Learning Mechanism – How does Aiva work?

Neural Network

Aiva uses CUDA, TITAN X Pascal GPUs, and cuDNN, along with TensorFlow deep learning algorithms. TensorFlow deep learning algorithms are programmed with reinforcement learning techniques.

First, deep learning maps inputs to outputs based on correlations (determining the relationship between two or more pieces of information) and approximations (matching data similar to predetermined data) using several layers of neural networks (set of algorithms). A neural network is modeled loosely after the human brain, which excels at pattern creation and recognition. Then, reinforcement learning (RL) is used.

Deep Leaning Mechanism

Neural Network

Reinforcement Learning

Reinforcement learning is a goal-oriented machine learning technique. RL algorithms learn how to reach a complex objective or maximize a dimension by gradually evolving from its own learning using the outputs occurring on every learning step starting from a blank state. Under the right conditions, they can achieve incredible performance. Say, for example, that you’re teaching AI technology to walk on a plain. While it’s performing a preliminary walk, it is also learning with every step it takes. When the programmer puts certain obstacles in its way, it may fall in the beginning, but from the fall it learns either how to dodge the obstructions or to walk over them. The AI technology evolves with the on-going walking and not only excels in plain terrain walking, but, when introduced to stairs, can also successfully climb stairs.

RL further allows machines and software agents to automatically determine the ideal behavior within a specific context in order to maximize its performance. Simple reward feedback is required for the agent to learn its behavior; this is known as the reinforcement signal. The algorithm does not require labeled inputs and outputs of data, consequently finding its own way around the data and improving its performance without being given any explicit instructions. This way, it is easier to capture the diversity and variation found in creative arts like music.

Reinforcement Learning

What did Aiva achieve?

Aiva learned to capture concepts of music theory and understand the art of music composition by reading through a large database of classical partitions written by famous composers. All of the partitions used to train Aiva are copyright-expired, which allows the company to continue training their AI free of cost. The Aiva team strategically chose to focus on classical music, as it is the style of music  most often used in movies, games, commercials, and trailer soundtracks. Once trained, AIVA composes its very own music, which is performed by professional artists on real instruments in a recording studio. Several Turing tests with music professionals have confirmed that the compositions can not be differentiated as human-produced or AI-produced

What is Aiva currently working on? How is Aiva being trained to learn unique styles of music? Will it be able to create its own signature style? Aiva’s CEO Pierre Barreau talked about Aiva’s ventures, at Con.Verge 2017. The talk will be released online soon. Stay tuned!


Leave a Reply

Your email address will not be published.

You may use these HTML tags and attributes:

<a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>