Novel Class Detection And Learning In Complex Scenarios

Machine learning technology has achieved remarkable success in numerous fields.Despite this,most current machine learning models are developed based on the closeworld assumption,meaning that the data environment is static and the models remain unchanged during the testing phase.However,in real-world scenarios,this simplified assumption is not applicable.The reason is that the real world is a complex,changing,and uncertain open environment.Faced with such situations,to ensure models can accurately classify known categories while also being capable of recognizing unknown items,discovering new categories,and gradually learning these new categories,open-world machine learning has become a focus of research.Although some progress has been made in this field,there are still many challenges in complex scenarios,including difficulty in distinguishing between categories,unknown of the number of categories,imbalance of category distribution,and scarcity of labeled data.To address these issues,this dissertation focuses on novel class detection and learning in complex scenarios.The main research content and innovations include the following aspects:1.For the problem of poor distinguishability between categories in open set recognition,this dissertation proposes an open set recognition algorithm based on Bayesian multiGaussian prototype learning.By constructing multi-Gaussian prototypes,the method more effectively represents the sample distribution of known classes and reserves adaptation space for unknown classes.Optimizing the model with both generative and discriminative constraints not only enhances intra-class compactness but also improves inter-class discrimination.The introduced Bayesian inference framework provides a solid theoretical basis for the model.Experimental results on different scenarios and datasets show that this method outperforms existing algorithms.In the unknown detection task,the AUROC score increased by 1-3%.Additionally,in the more complex open set recognition task,the F1 score increased by up to 10-12%.2.For the problem of difficulty of estimating the number of unknown categories in novel class discovery,this dissertation introduces a novel class discovery algorithm based on progressive bi-level contrastive learning.By conducting contrastive learning at both the prototype and prototype group levels,more accurate sample and prototype representations are obtained.An innovative prototype similarity measurement method is introduced,and by alternating between representation learning and prototype grouping,similar prototypes representing the same class are gradually gathered into the same prototype group,achieving novel class discovery with an unknown number of class.Experimental results demonstrate a more significant advantage over existing methods,with more accurate novel class identification and estimation of the number of novel class.In scenarios with sparse labels,the overall accuracy of novel class discovery improved by 7-9%.3.For the problem of class imbalance caused by the scarcity of novel class samples in new class discovery,this dissertation presents an imbalanced novel class discovery algorithm based on latent space community mining.Using a self-supervised learning approach to learn sample representations and prototypes representing the latent space distribution,a prototype network is constructed.The community division of prototypes is optimized using a modularity maximization objective function,revealing semantic class structures.To address class imbalance,a moving average-based strategy is introduced to estimate the prior of class imbalance.Experiments show that this method can effectively discover and identify imbalanced unknown novel classes in unlabeled data without prior class distribution information.Compared to existing methods,the algorithm improves accuracy in imbalanced novel class discovery tasks by an average of 9%.4.For the problem of labeled data scarcity in class-incremental learning,this dissertation proposes an unsupervised class incremental learning algorithm based on Gaussian mixture models.Utilizing fine-grained Gaussian mixture models and combining variational inference and optimal transport techniques,the algorithm models complex distributions in feature space.Gaussian parameters preserve historical information,effectively mitigating catastrophic forgetting.Unsupervised novel class discovery is achieved by maximizing the mutual information between the mixture Gaussian posterior and the class posterior.To further enhance recognition capability,additional constraint functions are introduced to reduce feature overlap between novel and old tasks.Experiments show that the method effectively discovers and learns novel classes in sequences of unlabeled tasks while avoiding the forgetting of historical class knowledge.Compared to existing methods,the algorithm improves overall accuracy by 7-10% and resistance to forgetting by 20-30%.