REGRESSOR - Build your neural network : Train your performance

[room]

http://www.audiobulb.com/REGRESSOR.htm

Audiobulb is absolutely thrilled to have teamed up with Jesse Stiles, a renowned composer, performer and educator to release Regressor.

Regressor is a Max for Live device that uses machine learning to map input signals (MIDI controllers, sensors, data streams) to parameters in Ableton Live. Instead of building control mappings by hand, you train Regressor with examples and it learns how to connect your inputs to any number of outputs. The result is a flexible, multidimensional control system you can set up in minutes.

Regressor works with a range of inputs making it a truly versatile performance interface:

- XY inputs

- MIDI inputs

- OSC inputs

How it works

Under the hood, Regressor uses a type of machine learning called regression. In simple terms, you show the device a set of examples: input signals paired with the output settings you want. It then learns how to interpolate between them. Once training is complete, your inputs control multiple parameters in Ableton simultaneously.

One key advantage: the number of inputs and outputs don't have to match. A single fader on a MIDI control surface (one dimension of input) can drive 16 different parameters on an effects rack. The neural network learns how to map that fader's movement across the full range of resulting sounds, with each parameter following its own non-linear output curve.

The reverse also works. A body-tracking system that produces 96 values (three coordinates for each of 32 body parts) would be impractical to map to sound parameters by hand. Regressor can take all 96 dimensions as input and let the neural network reduce them to a manageable number of outputs.

Manipulating outputs

Once Regressor is trained, two parameters let you shape the output signal. “Drift” (and its associated “speed” control) adds organic randomness to the predictions, a good way to introduce motion even when the input signal isn’t changing. “Slew” smooths the output, reducing jitter and slowing transitions. Both controls are only active after training and will be greyed out during training mode. The final performance can be captured via a "live" output mode, "recorded" and written into Ableton's automation lanes (making them editable after the fact) and listened back to via "playback" mode.