Automating composition: AI composer generates original melodies by ‘reading’ sheet music

In a world first, a deep-learning algorithm has been developed that generates original melodies in a given music style, without any knowledge of musical theory.

Developed by scientists at Ecole Polytechnique Fédérale de Lausanne (EPFL) in Switzerland, the algorithm, dubbed ‘Deep Artificial Composer’ or DAC, “trains” on sheet music of a given musical style before producing an original score in the same style of its own composition.

As a result, it does not produce audio files, but instead provides finished sheet music that humans can perform. At present DAC has been trained on Irish or Klezmer folk music, but it is designed to work with and mimic any musical style given to it.

“The deep artificial composer can produce complete melodies, with a beginning and an end, that are completely novel and that share features that we relate to style,” said EPFL scientist Florian Colombo, who developed DAC with the supervision of Wulfram Gerstner, director of the Computational Neuroscience Laboratory.

“To my knowledge, this is the first time that an artificial neural network model has produced entire and convincing melodies.”

Other AI composers have already been developed, however DAC is unique in that it does not use and implement musical theory. Instead it uses neural networks to determine probability distributions from existing melodies – that is, the likelihood of certain notes and note lengths occurring after each successive note.

DAV learns how music transitions from one note to the next, and the probability of both the pitch and the duration of the next note. Then it works through multiple scores, correcting its predictions and building up a model for that musical genre. Once it has successfully predicted 50% of successive notes and 80% of successive note durations, it is considered ‘trained’ and ready to compose its own music.

Working through note by note, DAC builds up an entire melody that is completely original, but is in the style of the music it has trained on. The result is melodies that sound completely man-made, but which are entirely produced by artificial intelligence.

Colombo playing a composition created by DAC on the cello. Images and music courtesy of EPFL

At present DAC is only able to produce melodies for one instrument or voice at a time, but Colombo hopes to in the future make it able to compose scores for entire orchestras in real-time.

This would realise an idea first proposed by mathematician Ada Lovelace in the 19th Century, but it will also likely cause fear among human composers, who join the pile of industries looking increasingly ready to be replaced by machines.

For others, though it could be hugely beneficial, bringing original music within reach of, for example, indie game developers and filmmakers.

The research was presented today at the Evostar conference in Amsterdam, the Netherlands.

Artificial intelligence would improve IVF’s success rate, research finds

Using artificial intelligence to determine embryos with the best chance of producing successful pregnancies through in-vitro fertilisation (IVF) would increase the procedure’s rate of success, according to a study.

The research, presented today at the 33rd Annual Meeting of the European Society of Reproduction and Embryology (ESHRE) in Geneva, found that using AI to standardise the selection of ‘good quality’ embryos would make viable embryo selection more accurate and so increase the chances of a successful pregnancy.

A popular choice for couples with fertility problems, IVF involves removing eggs from the woman’s ovaries, fertilising them with sperm to produce embryos and then implanting the most viable back into the womb, where they – all being well – will develop as normal into a healthy baby.

As part of this, embryologists determine which embryos are most viable for implantation, however despite this selection process, between 30 and 60% fail to implant successfully. Part of the reason for this is patient age, but there is also the fact that different embryologists will make different calls about whether an embryo is viable.

“The issue is that morphological grading by humans leads to wide inter and intra-operator variation,” said investigator Professor José Celso Rocha, from São Paulo State University, Brazil. In other words, embryologists’ assessment of an embryo’s shape and development results in significantly different conclusions depending on who is doing the assessing. And it is this variation that Rocha believes AI can help with.

“To classify images automatically will increase the predictive value of our embryo assessment,” he said. “By increasing objectivity and repeatability in embryo assessment, we can improve the accuracy of diagnosing embryo viability. Clinics can use this information as ‘artificial intelligence’ to customise treatment strategies and better predict a patient’s chance of pregnancy.”

The study which is the focus of this argument involved bovine embryos, with 482 seven-day-old embryos used to ‘train’ the AI system to recognise viable embryos from non-viable ones. The system assessed the embryos against 36 variables to determine viability, resulting in an accuracy of 76% – an improvement on conventional methods – and increased general consistency.

Now the research has moved onto human embryos, and although it is in the early stages of development, it is hoped that it will produce a highly repeatable system that will make future embryo classification far more consistent.

However, while AI might be poised to take over a part of the IVF process, it is human expertise that it will draw from.

“The artificial intelligence system must be based on learning from a human being,” said Rocha. “That is, the experienced embryologists who set the standards of assessment to train the system.”