Rolling element bearings are among the key components in many rotating machineries. It is hence necessary to determine the condition of the bearing with a reasonable degree of confidence. Many techniques have been developed for bearing fault detection. Each of these techniques has its own strengths and weaknesses. In this paper, various features are compared for detecting inner and outer race defects in rolling element bearings. Mutual information between the feature and the defect is used as a quantitative measure of quality. Various time, frequency, and time-frequency domain features are compared and ranked according to their cumulative mutual information content, and an optimal feature set is determined for bearing classification. The performance of this optimal feature set is evaluated using an artificial neural network with one hidden layer. An overall classification accuracy of 97% was obtained over a range of rotating speeds.