Efficient Learning of Linear Perceptrons