zakkbob

After reading over this (https://brilliant.org/wiki/backpropagation) many times I’ve finally got training working for the last layer. It shouldn’t be too hard to extend that to the rest.

I’m currently using a very naive approach of looping over data, so after training is working I’ll simplify the code a lot by making use of matrices.

I’ve attached the code for training a very simple network, it has one layer with one input and two outputs. It should’ve predicted [0.3, 0.25], so pretty close.

So far, I’ve implemented a feedforward neural network in go and written some tests.

Not sure what to put as an image, so here’s my ReLU function

I made it so nodes sync the blockchain on startup. Also, they exchange their public keys when they connect, I’ll use this to implement de-duplication of connections later! (I ran into a bug when implementing peer discovery where a node would just repeatedly connect to itself)

I know the demo isn’t very exciting but it shows the nodes connecting, exchanging keys, and syncing the blockchain.

Next is peer discovery!

Oh, I should also make the difficulty change dynamically and probably clean up the code a bit and maybe actually make it so transactions can be created through the cli, that seems important.