A semi-supervised deep-thinking concept based classifier for handwritten digit image recognition

Abstract

A new semi-supervised intelligent classifier based on human beings’ thinking process is developed. Unlike the learning processes in conventional classifiers, the complex mathematical optimization process is avoided. Instead, following the human beings’ thinking process, the design of the classifier is greatly simplified. In intuitive-imagery thinking layer, pattern clusters are initialized by the labelled samples and then trained via the unsupervised correlation principle. After that, a middle-level feature vector is generated for each pattern cluster via the dot-product operation. In the abstract-logical thinking layer, the moments’ information of the middle-level feature vectors is extracted to generate the advanced feature vector. The Bayesian inference based decision-making algorithm is then applied for classification. Like the human beings’ experience accumulation, a self-augmenting mechanism in the final layer is developed for adjusting the pattern clusters in real-time. An experiment for the classification of the handwritten digit images from the MNIST database is performed to show that the proposed intelligent pattern classifier can achieve high classification accuracy when the size of the labelled dataset is small.