Oscillating Neural Networks
The main idea behind SynthNN is to have a neural network running in real-time, while feeding forward its value to connected neurons. Each neuron in the network with multiple inputs will base its current value on the sum of the incoming waveforms. This will result in constructive and destructive interference.
On the left you can see a simple example of three neurons, x and y feed into z, the last one being the sum of the previous two.
The idea behind this is to experiment with a more fluid way of having floating point values interact with each other, and to try and avoid any type of backward propagation and gradient descent.
Using the oscillating values frequencies and amplitudes can be increased, and phase can be off-set to create a system that can be balanced.
Currently we use a very simple approach to balancing neurons, by just syncing the waveforms of neurons that tend to fire together to make their impact greater on the neurons down the line.
This effectively implements a rudimentary form of reinforcement learning.