One successful approach to feature extraction in face recognition problems is that of linear discriminant analysis (LDA). We examine an extension of this technique, called angular LDA, in which a non-linear transformation is applied after the LDA representation has been determined. We present experimental evidence, using the XM2VTS face database, that an ensemble of SVM classifiers operating in the angular LDA space is capable of making more accurate face verification and identification decisions than the same classifiers operating in the standard LDA space. We also compare experimentally the relative effectiveness of a number of techniques for ensemble design, ensemble decoding metric and SVM calibration algorithm.