An overview of algorithms for discovering new categories of problems based on unlabeled data

Abstract

In traditional machine learning, supervised and semi-supervised learning is designed to be used in a closed-world setting where the training data is fixed and does not change over time. Unfortunately, these methods still require a large number of labels for the categories to be categorized, which is expensive and impractical. A new category discovery algorithm is designed so that it can discover new categories while classifying and recognizing labeled images. In this case, the machine can automatically identify new categories without manual marking of image feature categories, which can greatly reduce the cost of image classification. Kai Han et al. named this problem a new category discovery problem and proposed that deep clustering can be used to solve it well. This paper focuses on the comparison of two commonly used robust baselines in the new category discovery and proposes that adding a post-processing model can better improve the accuracy of the model result. This paper applied the relaxed contrast learning method to the Ranking Statistics, and the accuracy of CIFAR-100 is improved by 6%.