On Warm-Starting Neural Network Training