Hybrid Neural Networks

After further consideration, we now have an implementation method for the hybrid neural networks.

First, we must reduce the size of the input to the initial convolutional network. This is necessary to reduce the size of the spiking network input. We initially attempted to train the spiking net with all 157 values (which expands to 629 neurons), but it failed to finish its first of 2000 training epochs after one hour, showing that it was utterly impractical.

We therefore used an FFT, which we clipped, to reduce the input size. This left us with 187 inputs to the traditional net, and so we chose the remaining layer sizes 45 -> 15 -> 5. This coded for 21 spiking input neurons, which even the slow algorithm can handle in a tractable length of time.

This very small spiking network input will not remain forever. We plan to expand it once we have a sufficiently efficient spiking neural network algorithm.