Towards Noise-adaptive, Problem-adaptive Stochastic Gradient Descent