<p>The task of Generalized Category Discovery (GCD) involves classifying unlabeled samples, which may belong to either seen or new classes, using a set of labeled samples from seen classes. The primary challenges stem from the absence of prior knowledge about the new classes and the uncertainty regarding the number of new classes. Three key problems need to be addressed: (1) Previous approaches often assumed the number of new classes to be known, which is not realistic. (2) The less discriminative representations for the new classes are a critical issue. (3) It is necessary to address the unreliable pseudo-labels assigned to the unlabeled samples. To address the GCD problem, we propose a method integrating three key ideas: (1) We propose a novel one-to-one pair matching approach to accurately estimate the number of new classes. (2) We develop an embedding network to project samples into a shared subspace, minimizing the distance between the new representations and their class centers, facilitating the learning of discriminative representations for both seen and new samples. (3) We propose a joint learning network that combines the embedding network and a classifier, leveraging neighborhood information to mitigate overconfidence in the pseudo-labels. Extensive experiments conducted on challenging datasets validate the effectiveness of our method. The experimental results demonstrate its capability for GCD.</p>