Continual Generalized Category Discovery (C-GCD) requires identifying novel classes from unlabeled data while retaining knowledge of known classes over time. Existing methods typically update classifier weights dynamically, resulting in forgetting and inconsistent feature alignment. We propose GOAL, a unified framework that introduces a fixed Equiangular Tight Frame (ETF) classifier to impose a consistent geometric structure throughout learning. GOAL conducts supervised alignment for labeled samples and confidence-guided alignment for novel samples, enabling stable integration of new classes without disrupting old ones. Experiments on four benchmarks show that GOAL outperforms prior methods, reducing forgetting by 16.1% and boosting novel class discovery by 3.2%, establishing a strong solution for long-horizon continual discovery.

This paper focuses on the Continual Test-Time Adaptation (CTTA) task, aiming to enable an agent to continuously adapt to evolving target domains while retaining previously acquired domain knowledge for effective reuse when those domains reappear. Existing shared-parameter paradigms struggle to balance adaptation and forgetting, leading to decreased efficiency and stability. To address this, we propose a frequency-aware shared and self-adaptive expert framework, consisting of two key components: (i) a dual-branch expert architecture that extracts general features and dynamically models domain-specific representations, effectively reducing cross-domain interference and repetitive learning cost; and (ii) an online Frequency-aware Domain Discriminator (FDD), which leverages the robustness of low-frequency image signals for online domain shift detection, guiding dynamic allocation of expert resources for more stable and realistic adaptation. Additionally, we introduce a Continual Repeated Shifts (CRS) benchmark to simulate periodic domain changes for more realistic evaluation. Experimental results show that our method consistently outperforms existing approaches on both classification and segmentation CTTA tasks under standard and CRS settings, with ablations and visualizations confirming its effectiveness and robustness.

Visual-language models (VLMs) like CLIP exhibit strong generalization but struggle with distribution shifts at test time. Existing training-free test-time adaptation (TTA) methods operate strictly within CLIP’s original feature space, relying on high-confidence samples while overlooking the potential of low-confidence ones. We propose MS-TTA, a training-free approach that enhances feature representations beyond CLIP’s space using a single-step k-nearest neighbors (kNN) Mean-Shift. By refining all test samples, MS-TTA improves feature compactness and class separability, leading to more stable adaptation. Additionally, a cache of refined embeddings further enhances inference by providing Mean-Shift-enhanced logits. Extensive evaluations on OOD and Cross-Dataset Benchmarks demonstrate that MS-TTA consistently outperforms state-of-the-art training-free TTA methods, achieving robust adaptation without requiring additional training.

Few-shot class-incremental Learning (FSCIL) enables models to learn new classes from limited data while retaining performance on previously learned classes. Traditional FSCIL methods often require fine-tuning parameters with limited new class data and suffer from a separation between learning new classes and utilizing old knowledge. Inspired by the analogical learning mechanisms of the human brain, we propose a novel analogical generative method. Our approach includes the Brain-Inspired Analogical Generator (BiAG), which derives new class weights from existing classes without parameter fine-tuning during incremental stages. BiAG consists of three components: Weight Self-Attention Module (WSA), Weight & Prototype Analogical Attention Module (WPAA), and Semantic Conversion Module (SCM). SCM uses Neural Collapse theory for semantic conversion, WSA supplements new class weights, and WPAA computes analogies to generate new class weights. Experiments on miniImageNet, CUB-200, and CIFAR-100 datasets demonstrate that our method achieves higher final and average accuracy compared to SOTA methods.

Domain-Incremental Learning (DIL) enables vision models to adapt to changing conditions in real-world environments while maintaining the knowledge acquired from previous domains. Given privacy concerns and training time, Rehearsal-Free DIL (RFDIL) is more practical. Inspired by the incremental cognitive process of the human brain, we design Dual-level Concept Prototypes (DualCP) for each class to address the conflict between learning new knowledge and retaining old knowledge in RFDIL. To construct DualCP, we propose a Concept Prototype Generator (CPG) that generates both coarse-grained and fine-grained prototypes for each class. Additionally, we introduce a Coarse-to-Fine calibrator (C2F) to align image features with DualCP. Finally, we propose a Dual Dot-Regression (DDR) loss function to optimize our C2F module. Extensive experiments on the DomainNet, CDDB, and CORe50 datasets demonstrate the effectiveness of our method.

Generalized Category Discovery (GCD) focuses on classifying known categories while simultaneously discovering novel categories from unlabeled data. However, previous GCD methods face challenges due to inconsistent optimization objectives and category confusion. This leads to feature overlap and ultimately hinders performance on novel categories. To address these issues, we propose the Neural Collapse-inspired Generalized Category Discovery (NC-GCD) framework. By pre-assigning and fixing Equiangular Tight Frame (ETF) prototypes, our method ensures an optimal geometric structure and a consistent optimization objective for both known and novel categories. We introduce a Consistent ETF Alignment Loss that unifies supervised and unsupervised ETF alignment and enhances category separability. Additionally, a Semantic Consistency Matcher (SCM) is designed to maintain stable and consistent label assignments across clustering iterations. Our method significantly enhancing novel category accuracy and demonstrating its effectiveness.

In dynamic real-world scenarios, continuous learning without forgetting old knowledge is essential, particularly in environments with stricter privacy protection or resource-constrained edge devices where storing old exemplars is infeasible. Therefore, Non-Exemplar Class-Incremental Learning (NECIL) has garnered significant attention. Compared with normal settings, it faces a more severe plasticity-stability dilemma and classifier bias. To address those challenges, we propose a framework based on the vision transformer architecture, called the Continual Expansion and Absorption Transformer (CEAT), which consists of two core components. First, we propose the Continual Expansion and Absorption (CEA) method to alleviate the trade-off between new and old classes by parallelly expanding a set of parameters (i.e. EF layer) on the backbone to learn new tasks, while freezing the backbone to retain old task knowledge. The EF layers can be seamlessly absorbed into the ViT backbone through parameter recombination before inference, mitigating storage and computational burdens. Second, we propose a Dynamic Boundary-Aware (DBA) method to generate dynamic pseudo-features for classifier calibration to address the classifier bias. Extensive experiments demonstrate that our approach achieves state-of-the-art performance, particularly showcasing significant improvements of 4.82% and 5.92% on TinyImageNet and ImageNet-Subset, respectively.

Chest X-ray (CXR) images have been widely adopted in clinical care and pathological diagnosis in recent years. Some advanced methods on CXR classification task achieve impressive performance by training the model statically. However, in the real clinical environment, the model needs to learn continually and this can be viewed as a domain incremental learning (DIL) problem. Due to large domain gaps, DIL is faced with catastrophic forgetting. Therefore, in this paper, we propose a Cross-domain invariant feature absorption and Domain-specific feature retention (CaD) framework. To be specific, we adopt a Cross-domain Invariant Feature Absorption (CIFA) module to learn the domain invariant knowledge and a Domain-Specific Feature Retention (DSFR) module to learn the domain-specific knowledge. The CIFA module contains the C(lass)-adapter and an absorbing strategy is used to fuse the common features among different domains. The DSFR module contains the D(omain)-adapter for each domain and it connects to the network in parallel independently to prevent forgetting. A multi-label contrastive loss (MLCL) is used in the training process and improves the class distinctiveness within each domain. We leverage publicly available large-scale datasets to simulate domain incremental learning scenarios, extensive experimental results substantiate the effectiveness of our proposed methods and it has reached state-of-the-art performance.

Online task-free continual learning (OTFCL) is a more challenging variant of continual learning that emphasizes the gradual shift of task boundaries and learning in an online mode. Existing methods rely on a memory buffer of old samples to prevent forgetting. However, the use of memory buffers not only raises privacy concerns but also hinders the efficient learning of new samples. To address this problem, we propose a novel framework called I$^{2}$CANSAY that gets rid of the dependence on memory buffers and efficiently learns the knowledge of new data from one-shot samples. Concretely, our framework comprises two main modules. Firstly, theInter-Class Analogical Augmentation(ICAN) module generates diverse pseudo-features for old classes based on the inter-class analogy of feature distributions for different new classes, serving as a substitute for the memory buffer. Secondly, theIntra-Class Significance Analysis(ISAY) module analyzes the significance of attributes for each class via its distribution standard deviation, and generates an importance vector as a correction bias for the linear classifier, thereby enhancing the capability of learning from new samples. We run our experiments on four popular image classification datasets: CoRe50, CIFAR-10, CIFAR-100, and CUB-200, our approach outperforms the prior state-of-the-art by a large margin.

Domain incremental object detection (DIOD) aims to gradually learn a unified object detection model from a dataset stream composed of different domains, achieving good performance in all encountered domains. The most critical obstacle to this goal is the catastrophic forgetting problem, where the performance of the model improves rapidly in new domains but deteriorates sharply in old ones after a few sessions. To address this problem, we propose a non-exemplar DIOD method named learning domain bias (LDB), which learns domain bias independently at each new session, avoiding saving examples from old domains. Concretely, a base model is first obtained through training during session 1. Then, LDB freezes the weights of the base model and trains individual domain bias for each new incoming domain, adapting the base model to the distribution of new domains. At test time, since the domain ID is unknown, we propose a domain selector based on nearest mean classifier (NMC), which selects the most appropriate domain bias for a test image. Extensive experimental evaluations on two series of datasets demonstrate the effectiveness of the proposed LDB method in achieving high accuracy on new and old domain datasets. The code is available at https://github.com/SONGX1997/LDB.

In this paper, we focus on a challenging Online Task-Free Class Incremental Learning (OTFCIL) problem. Dif-ferent from the existing methods that continuously learn the feature space from data streams, we propose a novel compute-and-align paradigm for the OTFCIL. It first com-putes an optimal geometry, i.e., the class prototype distri-bution, for classifying existing classes and updates it when new classes emerge, and then trains a DNN model by aligning its feature space to the optimal geometry. To this end, we develop a novel Dynamic Neural Collapse (DNC) algorithm to compute and update the optimal geometry. The DNC ex-pands the geometry when new classes emerge without loss of the geometry optimality and guarantees the drift distance of old class prototypes with an explicit upper bound. On this basis, we propose a novel DYnamic feature space Self-OrganizatioN (DYSON) method containing three ma-jor components, including 1) a feature extractor, 2) a Dy-namic Feature-Geometry Alignment (DFGA) module aligning the feature space to the optimal geometry computed by DNC and 3) a training-free class-incremental classifier de-rived from the DNC geometry. Experimental comparison results on four benchmark datasets, including CIFAR10, CI-FAR100, CUB200, and CoRe50, demonstrate the efficiency and superiority of the DYSON method. The source code is released at https://github.com/isCDX2IDYSON.

Despite the recent progress of class-incremental learning (CIL) methods, their capabilities in real-world scenarios such as multi-label settings remain unexplored. This paper focuses on a more practical CIL problem named multi-label class-incremental learning (MLCIL). MLCIL requires the vision models to overcome catastrophic forgetting of old knowledge while learning new classes from multi-label samples. Direct application of existing CIL methods to MLCIL leads to label absence, representative sample selection, and feature dilution problems. To address these problems, we present a novel AdaPtive Pseudo-Label-drivEn (APPLE) framework consisting of three components. First, the adaptive pseudo-label strategy is proposed to solve the label absence problem, which leverages the old model to annotate old classes for new samples. Second, a cluster sampling strategy is proposed to obtain more diverse samples to alleviate catastrophic forgetting under the MLCIL setting better. Finally, a class attention decoder is designed to mitigate the object feature dilution problem in multi-label samples. The extensive experiments on PASCAL VOC 2007 and MS-COCO demonstrate that our proposed method significantly outperforms other representative state-of-the-art CIL methods.

Object detection with the capacity to incrementally adapt to new domains is a crucial yet relatively under-explored research topic. The catastrophic forgetting problem presents a significant challenge to achieve this goal, where the model’s performance improves quickly in new conditions but deteriorates sharply in old ones after several incremental learning sessions. Drawing on recent discoveries in visual memories of the human brain, we introduce the Topology-Preserving Domain Incremental Object Detection (TP-DIOD) approach, which aims to address the catastrophic forgetting problem by extracting the topological structure of the feature space learned by the Convolutional Neural Network (CNN) model and preserving this topology during the subsequent incremental learning sessions. Specifically, we model the feature space topology using the self-organizing map (SOM) and construct an anchor image set based on the centroid vectors of the SOM nodes to memorize the feature space topology. We then develop the anchor loss function to penalize the topological changes of the feature space during the subsequent incremental learning sessions. Experimental evaluations on two sets of datasets demonstrate the effectiveness of the proposed TP-DIOD method in mitigating the catastrophic forgetting problem and achieving high accuracy on both old and new domain datasets.

FSCIL (Few-shot class-incremental learning) is a prominent research topic in the ML community. It faces two significant challenges: forgetting old class knowledge and overfitting to limited new class training examples. In this paper, we present a novel FSCIL approach inspired by the human brain’s analogical learning mechanism, which enables human beings to form knowledge about a target domain from the knowledge of the source domains that are analogical to the target in some aspects. The proposed analogical learning-based FSCIL (ALFSCIL) method consists of two major components: new class classifier constructor (NCCC) and Meta-Analogical training (MAT). The NCCC module utilizes a multi-head cross-attention transformer to compute analogies between new and old classes, generating new class classifiers by blending old class classifiers based on the computed analogies. The MAT module updates the parameters of the CNN feature extractor, the NCCC module, and the knowledge for each encountered class after each round of the FSCIL session. We turn the optimization process into a bi-level optimization problem(BOP) whose theoretical analysis proves the stability and plasticity of our proposed model. Experimental evaluations reveal that this proposed ALFSCIL method achieves the SOTA performance accuracies on three benchmark datasets: CIFAR100, miniImageNet, and CUB200.

Inspired by the global-local information processing mechanism in the human visual system, we propose a novel convolutional neural network (CNN) architecture named cognition-inspired network (CogNet) that consists of a global pathway, a local pathway, and a top-down modulator. We first use a common CNN block to form the local pathway that aims to extract fine local features of the input image. Then, we use a transformer encoder to form the global pathway to capture global structural and contextual information among local parts in the input image. Finally, we construct the learnable top-down modulator where fine local features of the local pathway are modulated by global representations of the global pathway. For ease of use, we encapsulate the dual-pathway computation and modulation process into a building block, called the global-local block (GL block), and a CogNet of any depth can be constructed by stacking a necessary number of GL blocks one after another. Extensive experimental evaluations have revealed that the proposed CogNets have achieved the state-of-the-art performance accuracies on all the six benchmark datasets and are very effective for overcoming the "texture bias" and the "semantic confusion" problems faced by many CNN models.

In this article, we focus on a new and challenging decentralized machine learning paradigm in which there are continuous inflows of data to be addressed and the data are stored in multiple repositories. We initiate the study of data-decentralized class-incremental learning (DCIL) by making the following contributions. First, we formulate the DCIL problem and develop the experimental protocol. Second, we introduce a paradigm to create a basic decentralized counterpart of typical (centralized) CIL approaches, and as a result, establish a benchmark for the DCIL study. Third, we further propose a decentralized composite knowledge incremental distillation (DCID) framework to transfer knowledge from historical models and multiple local sites to the general model continually. DCID consists of three main components, namely, local CIL, collaborated knowledge distillation (KD) among local models, and aggregated KD from local models to the general one. We comprehensively investigate our DCID framework by using a different implementation of the three components. Extensive experimental results demonstrate the effectiveness of our DCID framework. The source code of the baseline methods and the proposed DCIL is available at https://github.com/Vision-Intelligence-and-Robots-Group/DCIL.

In the context of incremental class learning, deep neural networks are prone to catastrophic forgetting, where the accuracy of old classes declines substantially as new knowledge is learned. While recent studies have sought to address this issue, most approaches suffer from either the stability-plasticity dilemma or excessive computational and parameter requirements. To tackle these challenges, we propose a novel framework, the Diverse Knowledge Transfer Transformer (DKT), which incorporates two knowledge transfer mechanisms that use attention mechanisms to transfer both task-specific and task-general knowledge to the current task, along with a duplex classifier to address the stability-plasticity dilemma. Additionally, we design a loss function that clusters similar categories and discriminates between old and new tasks in the feature space. The proposed method requires only a small number of extra parameters, which are negligible in comparison to the increasing number of tasks. We perform extensive experiments on CIFAR100, ImageNet100, and ImageNet1000 datasets, which demonstrate that our method outperforms other competitive methods and achieves state-of-the-art performance. Our source code is available at https://github.com/MIVXJTU/DKT.

Current class-incremental learning research mainly focuses on single-label classification tasks while multi-label class-incremental learning (MLCIL) with more practical application scenarios is rarely studied. Although there have been many anti-forgetting methods to solve the problem of catastrophic forgetting in single-label class-incremental learning, these methods have difficulty in solving the MLCIL problem due to label absence and information dilution problems. To solve these problems, we propose a Knowledge Restore and Transfer (KRT) framework containing two key components. First, a dynamic pseudo-label (DPL) module is proposed to solve the label absence problem by restoring the knowledge of old classes to the new data. Second, an incremental cross-attention (ICA) module is designed to maintain and transfer the old knowledge to solve the information dilution problem. Comprehensive experimental results on MS-COCO and PASCAL VOC datasets demonstrate the effectiveness of our method for improving recognition performance and mitigating forgetting on multi-label class-incremental learning tasks. The source code is available at https://gith.ub.com/witdsl/KRT-MLCIL.

Driven by practical needs, research on Class-Incremental Learning (CIL) has received more and more attentions in recent years. A technical challenge to be conquered by CIL methods is the catastrophic forgetting problem, where the model’s performance improves rapidly on new classes while deteriorates drastically on old ones. The main causes behind catastrophic forgetting include network drifts, inter-class confusions, etc. In this paper, we propose a novel CIL method that solves the catastrophic forgetting problem from two aspects. First, to solve the inter-class confusion problem, we propose a novel Semantic knOwledge gUided ciL framework (SOUL) that consists of a CNN feature extractor and a Bi-GCN (Graph Convolutional Network) classifier. In each CIL session, we use the semantic knowledge extracted from the class labels to build two inter-class relation graphs among all the encountered old and new classes. Using these two relation graphs, we develop a Bi-GCN classifier to fuse two kinds of semantic relations in a balanced way, and then to transfer the inter-class relations from semantic modality to image classification weights. The entire SOUL framework is trained end-to-end by the standard BP algorithm, which optimizes the Bi-GCN classifier and the CNN feature extractor jointly. Second, to prevent the network drift, we develop the local topology preserving strategy that divides the global topological structure of the learned feature space into a set of local topological relations, and maintains these local relations at CIL session. Experimental evaluations demonstrate the state-of-the-art performance accuracies on benchmark image classification datasets.

Deep models have shown to be vulnerable to catastrophic forgetting, a phenomenon that the recognition performance on old data degrades when a pre-trained model is fine-tuned on new data. Knowledge distillation (KD) is a popular incremental approach to alleviate catastrophic forgetting. However, it usually fixes the absolute values of neural responses for isolated historical instances, without considering the intrinsic structure of the responses by a convolutional neural network (CNN) model. To overcome this limitation, we recognize the importance of the global property of the whole instance set and treat it as a behavior characteristic of a CNN model relevant to model incremental learning. On this basis: 1) we design an instance neighborhood-preserving (INP) loss to maintain the order of pair-wise instance similarities of the old model in the feature space; 2) we devise a label priority-preserving (LPP) loss to preserve the label ranking lists within instance-wise label probability vectors in the output space; and 3) we introduce an efficient derivable ranking algorithm for calculating the two loss functions. Extensive experiments conducted on CIFAR100 and ImageNet show that our approach achieves the state-of-the-art performance.

Face Super-resolution (FSR) task works for generating high-resolution (HR) face images from the corresponding low-resolution (LR) inputs, which has received a lot of attentions because of the wide application prospects. However, due to the diversity of facial texture and the difficulty of reconstructing detailed content from degraded images, FSR technology is still far away from being solved. In this paper, we propose a novel and effective face super-resolution framework based on Transformer, namely Interconnected Dual Pyramid Transformer (IDPT). Instead of straightly stacking cascaded feature reconstruction blocks, the proposed IDPT designs the pyramid encoder/decoder Transformer architecture to extract coarse and detailed facial textures respectively, while the relationship between the dual pyramid Transformers is further explored by a bottom pyramid feature extractor. The pyramid encoder/decoder structure is devised to adapt various characteristics of textures in different spatial spaces hierarchically. A novel fusing modulation module is inserted in each spatial layer to guide the refinement of detailed texture by the corresponding coarse texture, while fusing the shallow-layer coarse feature and corresponding deep-layer detailed feature simultaneously. Extensive experiments and visualizations on various datasets demonstrate the superiority of the proposed method for face super-resolution tasks.

In this paper, we focus on the challenging few-shot class incremental learning (FSCIL) problem, which requires to transfer knowledge from old tasks to new ones and solves catastrophic forgetting. We propose the exemplar relation distillation incremental learning framework to balance the tasks of old-knowledge preserving and new-knowledge adaptation. First, we construct an exemplar relation graph to represent the knowledge learned by the original network and update gradually for new tasks learning. Then an exemplar relation loss function for discovering the relation knowledge between different classes is introduced to learn and transfer the structural information in relation graph. A large number of experiments demonstrate that relation knowledge does exist in the exemplars and our approach outperforms other state-of-the-art class-incremental learning methods on the CIFAR100, miniImageNet, and CUB200 datasets.

Deep models are vulnerable to catastrophic forgetting when fine-tuned on new data. Popular distillation-based methods usually neglect the relations between data samples and may eventually forget essential structural knowledge. To solve these shortcomings, we propose a structural graph knowledge distillation based incremental learning framework to preserve both the positions of samples and their relations. Firstly, a memory knowledge graph (MKG) is generated to fully characterize the structural knowledge of historical tasks. Secondly, we develop a graph interpolation mechanism to enrich the domain of knowledge and alleviate the inter-class sample imbalance issue. Thirdly, we introduce structural graph knowledge distillation to transfer the knowledge of historical tasks. Comprehensive experiments on three datasets validate the proposed method.

The ability to incrementally learn new classes is crucial to the development of real-world artificial intelligence systems. In this paper, we focus on a challenging but practical few-shot class-incremental learning (FSCIL) problem. FSCIL requires CNN models to incrementally learn new classes from very few labelled samples, without forgetting the previously learned ones. To address this problem, we represent the knowledge using a neural gas (NG) network, which can learn and preserve the topology of the feature manifold formed by different classes. On this basis, we propose the TOpology-Preserving knowledge InCrementer (TOPIC) framework. TOPIC mitigates the forgetting of the old classes by stabilizing NG's topology and improves the representation learning for few-shot new classes by growing and adapting NG to new training samples. Comprehensive experimental results demonstrate that our proposed method significantly outperforms other state-of-the-art class-incremental learning methods on CIFAR100, miniImageNet, and CUB200 datasets.