This repo is a very simple implementation of deep neural networks written in the Zig programming language.

It does not utilize tensors or any automatic differentiation library as it is only composed of fully connected feed forward layers.

To solve MNIST, create a new directory "data" in the root of the project, download the MNIST files from http://yann.lecun.com/exdb/mnist/, and unzip and store the files in the "data" directory. Then run:

zig build run

If the above url is not working for you, checkout: https://github.com/cvdfoundation/mnist.

here's some benchmark results from fiddling with hyper parameters etc spreadsheet

they're messy and look something like this:

feel free to give suggestions on namings etc in there, but it's mostly for personal tests and fiddling, perhaps eventually this whole project will be an actual library :)