SemiOccam is an efficient semi-supervised learning-based image recognition network using limited label data. To overcome the limitations of the complex structure and training process of existing methods, we build a hierarchical mixed density classification mechanism that optimizes the mutual information between feature representations and target classes, thereby removing unnecessary information and retaining important discriminant elements. Experimental results show that it achieves state-of-the-art performance on three common datasets, and achieves over 95% accuracy on two datasets using only four labeled samples per class. The simple structure allows the training time to be shortened to a few minutes. In addition, we uncovered the data leakage problem of the STL-10 dataset, which was overlooked in previous studies, and removed redundant data to obtain reliable experimental results, and released the refined CleanSTL-10 dataset.
