Character Recognition

2018

As my friend Johnny Lindbergh completed research on the use of neural networks in machine learning for his senior capstone project, we decided to experiment with building a simple neural network from the ground up--mostly to test our understanding of the math taking place in algorithms such as gradient descent.

The result of our labors is this network, a digit classifier capable of recognizing handwritten digits (from the MNIST dataset) with 89.1% accuracy, written from scratch.

This project deeply improved my understanding of both the C language and the calculus used in backpropagation. While there are certainly many improvements / optimizations that could be made to this network, it served its purpose of laying the foundations of my understanding of machine learning methods.

I also wrote a p5.js sketch (shown above) to take handwritten input, pass it through the net, and display the classification output vector. This demo can be found here.