Accelerated stochastic gradient descent via new model for first-order optimization