Implement a multilayer preceptron with C++.
Also a good excuse to build intuition on training by observing the MLP at each training step (dump out matplotlib figures of your weights and biases and make a movie out of them!)
Start with a horribly unoptimized, working prototype and explore performance gains through actual study and best practices. (Expand to profiling CUDA kernels once impl gets mature enough.)