CUDA Neural Net

A neural network for handwritten digit classification (MNIST) written from scratch in CUDA, with cuBLAS for the matrix maths. It's a small network (784 → 128 → 10, with ReLU and softmax) trained with SGD and momentum on the GPU, and it reaches 97.5% accuracy on the MNIST test set.

CUDA, cuBLAS, Deep Learning, C++

Code · All projects

By Rafay Ahmad (Bonevane).