Neural Network From Scratch

A 28×28 grayscale digit recognizer written from scratch in C++. Built while first learning how deep learning works, before I had even heard of PyTorch.

The thing that tripped me up wasn't the forward pass; it was that I'd written the gradient updates without actually understanding the chain rule properly, and ended up with models that "trained" (loss went down) but never generalized. Working through that confusion by hand (on paper, with shapes drawn out) is how I actually learned backpropagation. The code is the artifact; the real lesson lives in the margins of a notebook.

Don't follow this implementation too closely. It's the output of a learning sprint, not production code. But it served its purpose, which was building the intuition I now apply when reading PyTorch internals or writing kernels in Gradtuity.