Variance-scaled stochastic gradient descent

[accosmin]

• crazy idea: scale (minibatch-averaged) gradient with its variance: dx = avg(g) / (1 + var(g))
• use a simple decreasing learning rate schedule