Initialization-Dependent Sample Complexity of Linear Predictors and Neural Networks