Adaptive combinations of classifiers with application to on-line handwritten character recognition

Classifier combining is an effective way of improving classification performance. User adaptation is clearly another valid approach for improving performance in a user-dependent system, and even though adaptation is usually performed on the classifier level, also adaptive committees can be very effective. Adaptive committees have the distinct ability of performing adaptation without detailed knowledge of the classifiers. Adaptation can therefore be used even with classification systems that intrinsically are not suited for adaptation, whether that be due to lack of access to the workings of the classifier or simply a classification scheme not suitable for continuous learning. This thesis proposes methods for adaptive combination of classifiers in the setting of on-line handwritten character recognition. The focal part of the work introduces adaptive classifier combination schemes, of which the two most prominent ones are the Dynamically Expanding Context (DEC) committee and the ClassConfidence Critic Combining (CCCC) committee. Both have been shown to be capable of successful adaptation to the user in the task of on-line handwritten character recognition. Particularly the highly modular CCCC framework has shown impressive performance also in a doubly-adaptive setting of combining adaptive classifiers by using an adaptive committee. In support of this main topic of the thesis, some discussion on a methodology for deducing correct character labeling from user actions is presented. Proper labeling is paramount for effective adaptation, and deducing the labels from the user’s actions is necessary to perform adaptation transparently to the user. In that way, the user does not need to give explicit feedback on the correctness of the recognition results. Also, an overview is presented of adaptive classification methods for singleclassifier adaptation in handwritten character recognition developed at the Laboratory of Computer and Information Science of the Helsinki University of Technology, CIS-HCR. Classifiers based on the CIS-HCR system have been used in