Adding new classes. In the first experiment, we artificially created a situation of newly observed classes. First, we divided the pendigits data set into five sub sets S1,. .., S4 and T as shown in Table I (numbers correspond to occurences of samples per class). The sets S1,. .., S4 consist of samples from a growing number of classes, while T was used for testing. Then, we applied a number of incremental and online learning methods, where each time we started training on S1 and then increased the training set by the next sub set S2, S3, S4 be- fore re-training. This was done for two incremental learning methods, namely MSVDD [8] and MOCSVM [9], and four online learning algorithms, namely online Random Forests (ORF), online multi-class Gradient Boost (OMCGB), online multi-class LPBoost (OMCLP) [10]), and Mondrian forests (MF) [18]. The results are given in Table II. As we can TABLE II RESULTS FOR THE LEARNING SCENARIO OF TABLE I Data sets S1 S2 S3 S3 MSVDD 39.2% 58.68 % 79.21 % 98.23 % MOCSVM 39.8 % 59.25 % 77.95 % 95.18 % Mondrian Forests 39.21 % 58.78 % 78.36 % 95.21 % ORF 33.87 % 53.57 % 72.02 % 87.24 % OMCGB 29.48 % 34.67 % 59.28 % 60.03 % OMCLP 27.02 % 39.44 % 60.78 % 63.10 % see, incremental learning methods generally perform better, which is no surprise as they can rely on more exploitation of the training data. However, from the online methods the Mondrian forests clearly perform best. To increase the difficulty of the learning problem, we ran a second experiment, where only classes 0 and 1 were used for initial training. Then, in each learning round, we added Learning additional classes Influence of splitting
