T. Windeatt and R. Ghaderi. Binary labelling and decision level fusion. Information  Home/Search   Document Not in Database   Summary   Related Articles   Check

.... decomposition matrices are likely to have pairwise well separate codewords, that is high error recovering capabilities [7] Variants of the original ECOC algorithms have been proposed, as circular ECOC [21] in order to reduce the sensitivity to codeword selection, or binary labelling techniques [57] to reduce the correlation between the base learners. The good generalization properties of ECOC methods have been explained through the reduction of both bias and variance [32, 7] and by interpreting them as large margin classi ers [51, 2] Application of ECOC methods in several domains have ....

T. Windeatt and R. Ghaderi. Binary labelling and Decision Level Fusion. Information Fusion, 2(2):103-112, 2001. 41

....that a random code with a near equal column split of labels was theoretically better. Random codes were also shown in [17] to give Bayesian performance if pairs of code words were equidistant, and it was claimed that a long enough random code would not be outperformed by a pre defined code. In [33] a random assignment of class to codeword was suggested in order to reduce sensitivity to code word selection. Although various heuristics have been employed to produce better binary problemindependent codes there appears to be little evidence to suggest that performance significantly improves by ....

T. Windeatt and R. Ghaderi. Binary labelling and decision level fusion. Information

....Codes (ECOC) method being just one of them. In this paper we report on the novel use of ECOC for designing multiple experts for face verification. Use of ECOC for decomposing a multi class problem into a set of complementary two class problems is a well established method in many applications [4, 5, 6, 7, 11, 21, 23, 22, 24]. Such a decomposition means that attention can be focused on developing an effective technique for the two class classifier, without having to consider explicitly the design and automation of the multi class case. It is also hoped that the parameters of a simple expert run many times may be ....

....columns be high, in order to ensure diversity among experts. In our experiments we use BCH coding method with IKJ:JML N =PO code word removed, and we obtain equidistant rows by the over produce and select strategy. Further details, and a comparison of BCH with random codes can be found in [21]. From another perspective, we can view equation 1 as providing posterior probability of super class membership. The decomposition defined by can then be interpreted as a transformation between spaces representing probabilities of individual class and super class membership [24] In fact, a ....

T. Windeatt and R. Ghaderi. Binary labelling and decision level fusion. Information Fusion, 2(2):103--112, 2001.

....that classi cation performance with LS can be comparable to the original distance based approach. 1 Introduction Use of Error Correcting Output Codes (ECOC) for decomposing a multi class problem into a set of complementary two class problems is a well established method in many applications [1, 2, 4 13, 15 18]. When rst suggested ECOC was based on the idea of using error correcting codes as class labels, so that individual classi cation errors propagated from a set of binary classi ers can potentially be corrected [4] For a two class problem, classi cation errors can be one of two types, either ....

T. Windeatt and R. Ghaderi. Binary labelling and decision level fusion. Information fusion, to be published, 2001.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC