This module contains an artificial neural network. An artificial intelligence that is completely untrained. The type of the artificial neural net is called perceptron. The idea of this module is not to create breathtaking melodies and rhythms on command, as one might think. We dig a little deeper and focus on the learning process itself. How do machines learn and what does it sound like when you hear them learn? What do they sound like in the imperfect early stages of learning and which paths does the progress of learning take? Our module brain provides the opportunity explore the sonic developments a learning AI systems take.
Example: The following example tries to demonstrate the effects of the very obvious time-based use case: A melody is a sequence of notes ordered by time. This means: two parameters, we call them X and Y. At each point in time X there is a value for the pitch Y. The Xin and Yin sockets serve as inputs. Connect Yin to the pitch CV of you melody source. Xin should be connected to another CV representing the corresponding point in time. So every point in time of the repeating melody, Xin should be a corresponding unique value. For Xin a ascending or descending voltage ramp – which can be easily created using a slow LFO sawtooth which synchronously restarts with the melody – could be used. In such a setup the Yout should start to imitate what was provided at Yin. This is what the artificial net learned so far about the melody. After a certain time the net should more or lesse perfectly repeat the melody.
This is the most simple way of using the module and listening to the progress of the net can be interesting. But imaging if you put two different CVs in relation which are not exactly time-based. Or feeding audio signals into the net. But keep in mind that for technical reasons the module is only capable of working safely with low audio frequencies – what will happen if we try higher ones?
How to Use
Input Jack xin: reference voltage for the training voltages.
Input Jack yin: training voltages (to be learned).
Output Jack yout: output voltage of the neural network (what has been learned so far)
Knobs scale and shift: calibrate the output voltage (see below)
Output jack and LED error: a value between 0 and 5V representing the current level of error.
Button and jack reset: reset the neural network into an untrained state.
Switch freeze: stop learning but keep processing input and output voltages.
Knob A or α: set the learning rate of the neurons.
Knob B or β: set the learning rate of the entire network.
Technical Details
Calibration
The module already comes calibrated (knobs scale and shift in middle position) but over time a recalibration could become necessary.
In order to verify output voltages: connect your yin signal to the yin input jack and to an oscilloscope using a split cable. Connect the yout to another channel of your oscilloscope. Make sure both channels are displayed with the same configuration. Now turn the knobs shift and scale until both cuves perfectly match each other.
If the limitation of shift and scale are not enough it is very likely that the module is broken and it may be necessary to send the module in forr service. However: if the module is removed from the rack you have access to the trimmer potentiometers RV2 (shift) and RV1002 (scale) which may be used to pre-calibrate the module. Turn the knobs scale and shift into middle position and use the trimmer potentiometers to calibrate the output.
Changing / Uploading Software
If you ever want to change/update the software, here is the process:
- Remove the module from your rack.
- Close the “BOOTP” jumper.
- Connect the module to your computer via USB.
- A new Drive called “RPI-RP2” will appear.
- Drag/copy the new software file (.uf2) onto the drive.
- The “RPI-RP2” drive will auto-disconnect and reconnect from/to your computer. Ignore if the computer complains about unsafe removal of an external drive.
- Disconnect the module from your computer.
- Open the “BOOTP” jumper
- Put the module back into your rack.
Known limitations
- Depending on the configuration of the net a delay between input and output of very few milliseconds will occur. This should be no problem for common CV use cases.
- Due to technical reasons the module will work with voltage levels between -8V and +8V.