I've developed a model that incorporates STDP and neuromodulation under a single computational unit. A network of the neurons like these solves MNIST with ~96% accuracy without backprop in under 5 epochs of unsupervised feature extraction, or ~92% after just 1 epoch. I have integrated it into pytorch to use traditional benchmarking methods locally.

A GitHub repo with the non-pytorch implementation is available, along with tooling - https://github.com/arterialist/neuron-model