We present some recent results with improved convergence rates for gradient descent on matrix factorization problems, made possible by using an unbalanced initialization of the weights. Applications to neural networks and fine-tuning will be discussed.