[Neural Net {Python}] (#pull) Introduction simple feedforward gradient descent with backprop (not vectorized, not optimized) accuracy: 95% (9516/10000)