Can computers create art?

AlgorRitmo is an experiment on music creation by a hybrid human-artificial intelligence

Computers, long believed to be good only at crunching numbers, are making fast strides towards solving complex cognitive challenges such as driving cars, maintaining a conversation, and even being better-than-humans at playing video games as well as once-considered-too-difficult board games like Go.

As people apply these new advances to content creation (text, images, and music) they face not only the usual technical challenges, but also the fact that the quality of music, and art in general, is very subjective: the same piece can loose popularity in a very short time, while others need to be remade with current styles to appeal to younger audiences.

Next

The machine needs to aquire good taste

...and it cannot do it alone.

The field of machine learning is defined by the datasets from which the algorithms can extract useful information. An artificial intelligence could learn how to compose from reading vast music libraries, but it would never get out of what is already known. In order to surpass us and explore approaches a human would never try, the algorithm needs feedback on what we like and what we don't. AlgorRitmo will be at the Sonar music festival, where the audience will be able to vote on the songs created by the computer, as well as create original pieces of music on the spot and tell us wether what they made sounds good or bad. AlgorRitmo will run live from Sonar on the Minotauro supercomputer of the Barcelona Supercomputing Center

Next

How does AlgorRitmo work?

The devil is in the details

The devil is in the details At its core, AlgorRitmo tries to predict the most likely note that needs to be played after a given sequence. This is done using a type of algorithm called recurrent neural networks. During the festival, the audience will vote if songs are good or not, and this data will flow back to AlgorRitmo to influence the next batch of generated songs. One last, but not least important, part of the project is the synthetization of sound. AlgorRitmo will produce musical scores (computer readable music or MIDI files) that can be used for inspiration or accompanying tracks using any instrument of choice, as a guest musician will do during the festival. However, we will also generate music automatically using a set of instruments created from samples of the electronic sound of Marenostrum

Next

AlgorRitmo uses the following technologies

We are deeply thankful to all those that facilitated us access to them

GPU computing

Neural networks can be trained real fast using special accelerated hardware.

Neural network libraries

Mainly Keras, Theano, and Tensorflow.

Do you want to know more?

Visit our blog