Neural Networks from Scratch
The original Python project from 2017 was an attempt to understand what at the time felt like new technology. Today, we are of course keenly aware of machine-learning based technology, as am I now keenly aware that neural networks originated about 50 years before I was born.
My goal was to implement the network and train it on the MNIST handwritten digit dataset, without the use of any machine learning libraries. In the end I achieved a result of about 95% accuracy. It is happily/unfortunately the most advanced math I have had to do in a programming project yet.
Years later, in 2023, I used the original Python code as a reference to implement a similar neural network in Rust. As you might expect, the Rust version ran quite a bit faster, which allowed me to iterate more on tuning the network, ending up with 98.5% accuracy on the same dataset. It also gave me many thoughts about Rust, none of which are informed enough to share here.