Early stages of writing up about my project to create a neural network in c++. over time i want to improve training latency of the net via SW optimisations. But also it serves as a foundation to explore CUDA processing to make incredibly fast training loops - or atleast compared to cpu?

Github link: @codegen-cnn

Will create a series on this - currently the training algorithm is taking a while. I think I need to do vectorised updates for matrix processing with SIMD? Currently i’m doing the naive “update each cell in the matrix on its own” :/