Êêêê Blockinøøóò Ññøøóó× Óö Ò Ôøøú Óñññøøøø Ðð×××¢¢ö

Matti Aksela1;2, Jorma Laaksonen1, Erkki Oja1 Helsinki University of Te hnology Neural Networks Resear h Centre P.O.Box 5400, Fin-02015 HUT, Finland fmatti.aksela,jorma.laaksonen,erkki.ojag hut. Jari Kangas Nokia Resear h Center P.O.Box 100, Fin-33721 Tampere, Finland jari.a.kangas nokia. om Abstra t Adaptation is an effe tive method for improving lassi ation a ura y and a ommittee stru ture an in general improve on its members' performan e. Therefore an adaptive ommittee stru ture is a tempting approa h. Reje tion may be used in handwriting re ognition to improve performan e through either dire ting the problemati hara ter to a spe ial lassi er that handles su h hard ases or disarding it. The experiments in this paper ompare several fundamentally different approa hes to implementing reje tion in an adaptive ommittee lassi er. A Dynami ally Expanding Context (DEC) based ommittee is used for evaluating these approa hes. The results show that if the reje ted lasses are handled with a 50% error rate, the performan e is improved. A s heme in whi h there is an adjustable threshold for distan e-based reje tion is an effe tive method for implementing reje tion in this setting. 1 Introdu tion In lassi ation it is a ommon approa h to use a set of referen e samples to mat h the input sample against. If there is signi ant variation, having a large enough number of referen e samples may qui kly be ome impra ti al or even impossible. In su h ases lassi er adaptation is an effe tive method for improving performan e. Another approa h to improve re ognition performan e is to ombine different lassi ers in a ommittee. This is feasible be ause in the outputs of several lassi ers the errors are not ne essarily overlapping. Although the most ommon way of adaptation is to adapt a single re ognizer, it 1A knowledgement: This resear h was partly nan ed by the proje t New Information Pro essing Prin iples, Finnish Centre of Ex ellen e Programme 2000-2005, A ademy of Finland. 2A knowledgement: This resear h was partly nan ed by a grant from the Nokia Foundation. 2nd Classifier Incorrect Correct