Weights with the arcs showing which notes in the past are informing the future. Primarily this involves developing new deep learning and reinforcement learning algorithms for generating songs, images, drawings, and other materials. To see the self-reference, we visualized the last layer of attention Magenta is a research project exploring the role of machine learning in the process of creating art and music. Grayed out blocks), culminating with a quick succession to build tension. Nevertheless, a lot of work remains before Magenta models are writing. These models improve on the initial Magenta Basic RNN by adding two forms of memory manipulation, simple lookback and learned attention. Motif (identifiable through the denser sections with broken lines in the opening visualization), then repeats and varies it several times in the piece (manually marked by To train the Attention RNN on your own MIDI collection and generate your own melodies from it, follow the steps in the README on GitHub. In the following example, the model introduces a rhythmically quirky tremolo New algorithm for relative self-attention thatĭramatically reduces the memory footprint, allowing us to scale to musical The previous relative attention paper used an algorithm that was overly Which is not possible with the original Transformer model. This setup enables you to generate AI music based on melodies played in real-time. The coolest part of the project is interacting with the model utilizing Magenta’s midi interface in Ableton. Relative self-attentionĪlso allows the model to generalize beyond the length of the training examples, The trained models are capable of generating monophonic melodies given a primer melody. The model is able to focus more on relational features. Which explicitly modulates attention based on how far apart two tokens are, We found that by using relative attention, Track of regularity that is based on relative distances, event orderings, and periodicity. While the original Transformer allows us to capture self-reference throughĪttention, it relies on absolute timing signals and thus has a hard time keeping Our recent Wave2Midi2Wave project also uses Music Transformer-based model that has direct access to all earlier
In contrast to an LSTM-based model like Performance RNN thatĬompresses earlier events into a fixed-size hidden state, here we use a Google researcher Douglas Eck showing off an AI-generated image at Moogfest. That allows us to generate expressive performances directly (i.e. Google is launching a new research project to see if computers can be truly creative. Similar to Performance RNN, we use an event-based representation