Fuzzy learning algorithms provide an efficient method for generating approximating functions from training data. The approximations are obtained by constructing a fuzzy associative memory whose entries are determined locally using the information contained in the training set. A two-level fuzzy learning algorithm is introduced that incorporates error analysis into the approximation. The addition of a second fuzzy associative memory provides a significant improvement in the accuracy of the resulting approximations. The construction requires only one additional pass through the training data, maintaining the efficiency of the fuzzy learning algorithms. The improvement in accuracy is obtained without increasing the number of training instances, making this technique particularly suitable for problem domains in which the training information is unavailable or expensive to obtain.