Inspiration
Interest in the dynamics of neural network weight updates in supervised classification
What it does
Maps weights to frequencies within a 7-octave spectrum, and allows the user to modulate the periods of those frequencies by placing their hand a certain distance from a sensor. Periods decrease closer to the sensor, and increase farther away. A servo and a can are used as a makeshift drumset.
How we built it
Arduino neural net, made with the NeuralDuino link library. Functions that update and print weights were modified to play a tone and modulate it with the sensor.
Challenges we ran into
We initially set out to implement a Physical Neural Net, in circuitry and software. We found that we did not have the tools or knowledge of physics to fulfill this goal, so we used the existing model we had built to make music.
Accomplishments that we're proud of
How dynamic the sensor modulation is - it's essentially a theremin.
What we learned
A lot of the physics and math of electromagnetism
What's next for MuNet
All sorts of modulation (the goal is a real synthesizer), and making the net create music by more than just updating weights (initialization of activation function, etc).
Built With
- ai
- arduino
- c
- c++
- iot
- linux
- machine-learning
- robot
Log in or sign up for Devpost to join the conversation.