Active Learning in Approximately Linear Regression Based on Conditional Expectation of Generalization Error