About Retraining Rule in Multi-Expert Intelligent System for Semi-Supervised learning using SVM classifiers

Training a system for pattern recognition is a task that require a large amount of labeled data. However, the creation of such training set is often difficult, expensive and time consuming because it requires the efforts of experienced human annotators. On the other hand, unlabeled data may be relatively easy to collect, but there are few ways to use them. Semi-Supervised learning is a useful approach to reduce human labor and improve accuracy using unlabeled data, together with labeled data. This paper proposes three methods in order to re-train classifiers in a multi-expert scenario, when new (unknown) data are available. In fact, when a multi-expert system is adopted, the collective behavior of classifiers can be used both for recognition aims and also selection of the most profitable samples for system re-train. More specifically a misclassified sample for a particular expert can be used to update the expert itself if the collective behavior of the multi-expert system allows to classify the sample with high confidence. In addition, this paper provides a comparison between the new approach and those available in literature for semi-supervised learning using the SVM classifier by taking into account four different combination techniques at abstract and measurement level. The experimental results, that have been obtained using the handwritten digits of the CEDAR database, demonstrate the effectiveness of the proposed approach.