Hogwild: A Lock-Free Approach to Parallelizing Stochastic Gradient Descent