English

Abstract

Abstract This report points out the role of sequences of samples for training an incremental learn-ing method. We de ne characteristics of incremental learning methods to describe theinuence of sample ordering on the performance of a learned model. Di erent types ofexperiments evaluate these properties for two di erent datasets and two di erent incre-mental learning methods. We show how to nd sequences of classes for training just basedon the data to get always best possible error rates. This is based on the estimation ofBayes error bounds. 1.1 Introduction In this section the goal of learning is identifying and discriminating instances of classes.This requires generative models as they are most descriptive and assure learning of manyclasses. On the other hand one needs discriminative models as they are more ecientin separating classes. We will use both of them, either alone or in combination to useproperties of both models.Incremental learning is clearly in the scope of eTRIMS. It is impossible to capturethe huge variability of facades with one dataset. Instead we need a continuous learningsystem that is able to improve already learned models using new examples. This processin principle does not stop.For this purpose there exists a number of incremental learning methods. But one caneasily show that the performance of these methods depends on the order of examples fortraining, also this is not mentioned by most other authors.Imagine little kids, they are sophisticated incremental learners. They permanentlyincrease their knowledge just by observing their surrounding. But obviously they learnmuch better when examples are presented in a meaningful order.Therefore teachers design a curriculum for teaching their pupils in the beginning of aschool year. This way they ensure that the topics are presented to the pupils in a wellstructured order depending of the complexity and challenge of each single topic.Analogously, we want to de ne curricula for training classi cation procedures usingincremental learning methods. We want to de ne good sequences of examples to trainsuch methods such that they always perform best.For this, we rst characterize incremental learning methods regarding the e ects ofsample ordering in Section 1.2. Thereafter we propose experiments to explore these e ectsand show results for two di erent types of incremental learning methods and two di erenttypes of data in Section 1.3. Finally we propose a method to de ne suitable class sequencesfor training based on the estimate of bounds of the Bayes error in Section 1.4.