Vision Language Models (VLMs), pre-trained on large-scale image-text datasets, enable zero-shot predictions for unseen data but may underperform on specific unseen tasks. Continual learning (CL) can help VLMs effectively adapt to new data distributions without joint training, but faces challenges of catastrophic forgetting and generalization forgetting. Although significant progress has been achieved by distillation-based methods, they exhibit two severe limitations. One is the popularly adopted single-teacher paradigm fails to impart comprehensive knowledge, The other is the existing methods inadequately leverage the multimodal information in the original training dataset, instead they rely on additional data for distillation, which increases computational and storage overhead. To mitigate both limitations, by drawing on Knowledge Integration Theory (KIT), we propose a Multi-Stage Knowledge Integration network (MulKI) to emulate the human learning process in distillation methods. MulKI achieves this through four stages, including Eliciting Ideas, Adding New Ideas, Distinguishing Ideas, and Making Connections. During the four stages, we first leverage prototypes to align across modalities, eliciting cross-modal knowledge, then adding new knowledge by constructing fine-grained intra- and inter-modality relationships with prototypes. After that, knowledge from two teacher models is adaptively distinguished and re-weighted. Finally, we connect between models from intra- and inter-task, integrating preceding and new knowledge. Our method demonstrates significant improvements in maintaining zero-shot capabilities while supporting continual learning across diverse downstream tasks, showcasing its potential in adapting VLMs to evolving data distributions.

Generalized Category Discovery (GCD) aims to classify both base and novel images using labeled base data. However, current approaches inadequately address the intrinsic optimization of the co-occurrence matrix $\bar{A}$ based on cosine similarity, failing to achieve zero base-novel regions and adequate sparsity in base and novel domains. To address these deficiencies, we propose a Non-Negative Generalized Category Discovery (NN-GCD) framework. It employs Symmetric Non-negative Matrix Factorization (SNMF) as a mathematical medium to prove the equivalence of optimal K-means with optimal SNMF, and the equivalence of SNMF solver with non-negative contrastive learning (NCL) optimization. Utilizing these theoretical equivalences, it reframes the optimization of $\bar{A}$ and K-means clustering as an NCL optimization problem. Moreover, to satisfy the non-negative constraints and make a GCD model converge to a near-optimal region, we propose a GELU activation function and an NMF NCE loss. To transition $\bar{A}$ from a suboptimal state to the desired $\bar{A}^*$, we introduce a hybrid sparse regularization approach to impose sparsity constraints. Experimental results show NN-GCD outperforms state-of-the-art methods on GCD benchmarks, achieving an average accuracy of 66.1\% on the Semantic Shift Benchmark, surpassing prior counterparts by 4.7\%.

Video Wire Inpainting (VWI) is a prominent application in video inpainting, aimed at flawlessly removing wires in films or TV series, offering significant time and labor savings compared to manual frame-by-frame removal. However, wire removal poses greater challenges due to the wires being longer and slimmer than objects typically targeted in general video inpainting tasks, and often intersecting with people and background objects irregularly, which adds complexity to the inpainting process. Recognizing the limitations posed by existing video wire datasets, which are characterized by their small size, poor quality, and limited variety of scenes, we introduce a new VWI dataset with a novel mask generation strategy, namely Wire Removal Video Dataset 2 (WRV2) and Pseudo Wire-Shaped (PWS) Masks. WRV2 dataset comprises over 4,000 videos with an average length of 80 frames, designed to facilitate the development and efficacy of inpainting models. Building upon this, our research proposes the Redundancy-Aware Transformer (Raformer) method that addresses the unique challenges of wire removal in video inpainting. Unlike conventional approaches that indiscriminately process all frame patches, Raformer employs a novel strategy to selectively bypass redundant parts, such as static background segments devoid of valuable information for inpainting. At the core of Raformer is the Redundancy-Aware Attention (RAA) module, which isolates and accentuates essential content through a coarse-grained, window-based attention mechanism. This is complemented by a Soft Feature Alignment (SFA) module, which refines these features and achieves end-to-end feature alignment. Extensive experiments on both the traditional video inpainting datasets and our proposed WRV2 dataset demonstrate that Raformer outperforms other state-of-the-art methods.

While anti-amnesia FSCIL learners often excel in incremental sessions, they tend to prioritize mitigating knowledge attrition over harnessing the model's potential for knowledge acquisition. In this paper, we delve into the foundations of model generalization in FSCIL through the lens of the Neural Tangent Kernel (NTK). Our primary design focus revolves around ensuring optimal NTK convergence and NTK-related generalization error, serving as the theoretical bedrock for exceptional generalization. To attain globally optimal NTK convergence, we employ a meta-learning mechanism grounded in mathematical principles to guide the optimization process within an expanded network. Furthermore, to reduce the NTK-related generalization error, we commence from the foundational level, optimizing the relevant factors constituting its generalization loss. Specifically, we initiate self-supervised pre-training on the base session to shape the initial network weights. Then they are carefully refined through curricular alignment, followed by the application of dual NTK regularization tailored specifically for both convolutional and linear layers. Through the combined effects of these measures, our network acquires robust NTK properties, significantly enhancing its foundational generalization. On popular FSCIL benchmark datasets, our NTK-FSCIL surpasses contemporary state-of-the-art approaches, elevating end-session accuracy by 2.9% to 8.7%.

Few-shot Class-Incremental Learning (FSCIL) aims at learning new concepts continually with only a few samples, which is prone to suffer the catastrophic forgetting and overfitting problems. The inaccessibility of old classes and the scarcity of the novel samples make it formidable to realize the trade-off between retaining old knowledge and learning novel concepts. Inspired by that different models memorize different knowledge when learning novel concepts, we propose a Memorizing Complementation Network (MCNet) to ensemble multiple models that complements the different memorized knowledge with each other in novel tasks. Additionally, to update the model with few novel samples, we develop a Prototype Smoothing Hard-mining Triplet (PSHT) loss to push the novel samples away from not only each other in current task but also the old distribution. Extensive experiments on three benchmark datasets, e.g., CIFAR100, miniImageNet and CUB200, have demonstrated the superiority of our proposed method.

The domain shift between the source and target domain is the main challenge in Cross-Domain Few-Shot Learning (CD-FSL). However, the target domain is absolutely unknown during the training on the source domain, which results in lacking directed guidance for target tasks. We observe that since there are similar backgrounds in target domains, it can apply self-labeled samples as prior tasks to transfer knowledge onto target tasks. To this end, we propose a task-expansion-decomposition framework for CD-FSL, called Self-Taught (ST) approach, which alleviates the problem of non-target guidance by constructing task-oriented metric spaces. Specifically, Weakly Supervised Object Localization (WSOL) and self-supervised technologies are employed to enrich task-oriented samples by exchanging and rotating the discriminative regions, which generates a more abundant task set. Then these tasks are decomposed into several tasks to finish the task of few-shot recognition and rotation classification. It helps to transfer the source knowledge onto the target tasks and focus on discriminative regions. We conduct extensive experiments under the cross-domain setting including 8 target domains: CUB, Cars, Places, Plantae, CropDieases, EuroSAT, ISIC, and ChestX. Experimental results demonstrate that the proposed ST approach is applicable to various metric-based models, and provides promising improvements in CD-FSL.

Semantic information provides intra-class consistency and inter-class discriminability beyond visual concepts, which has been employed in Few-Shot Learning (FSL) to achieve further gains. However, semantic information is only available for labeled samples but absent for unlabeled samples, in which the embeddings are rectified unilaterally by guiding the few labeled samples with semantics. Therefore, it is inevitable to bring a cross-modal bias between semantic-guided samples and nonsemantic-guided samples, which results in an information asymmetry problem. To address this problem, we propose a Modal-Alternating Propagation Network (MAP-Net) to supplement the absent semantic information of unlabeled samples, which builds information symmetry among all samples in both visual and semantic modalities. Specifically, the MAP-Net transfers the neighbor information by the graph propagation to generate the pseudo-semantics for unlabeled samples guided by the completed visual relationships and rectify the feature embeddings. In addition, due to the large discrepancy between visual and semantic modalities, we design a Relation Guidance (RG) strategy to guide the visual relation vectors via semantics so that the propagated information is more beneficial. Extensive experimental results on three semantic-labeled datasets, i.e., Caltech-UCSD-Birds 200-2011, SUN Attribute Database, and Oxford 102 Flower, have demonstrated that our proposed method achieves promising performance and outperforms the state-of-the-art approaches, which indicates the necessity of information symmetry.

Zero-Shot Learning (ZSL) is generally achieved via aligning the semantic relationships between the visual features and the corresponding class semantic descriptions. However, using the global features to represent fine-grained images may lead to sub-optimal results since they neglect the discriminative differences of local regions. Besides, different regions contain distinct discriminative information. The important regions should contribute more to the prediction. To this end, we propose a novel stacked semantics-guided attention (S2GA) model to obtain semantic relevant features by using individual class semantic features to progressively guide the visual features to generate an attention map for weighting the importance of different local regions. Feeding both the integrated visual features and the class semantic features into a multi-class classification architecture, the proposed framework can be trained end-to-end. Extensive experimental results on CUB and NABird datasets show that the proposed approach has a consistent improvement on both fine-grained zero-shot classification and retrieval tasks.

Zero-Shot Learning (ZSL) is generally achieved via aligning the semantic relationships between the visual features and the corresponding class semantic descriptions. However, using the global features to represent fine-grained images may lead to sub-optimal results since they neglect the discriminative differences of local regions. Besides, different regions contain distinct discriminative information. The important regions should contribute more to the prediction. To this end, we propose a novel stacked semantics-guided attention (S2GA) model to obtain semantic relevant features by using individual class semantic features to progressively guide the visual features to generate an attention map for weighting the importance of different local regions. Feeding both the integrated visual features and the class semantic features into a multi-class classification architecture, the proposed framework can be trained end-to-end. Extensive experimental results on CUB and NABird datasets show that the proposed approach has a consistent improvement on both fine-grained zero-shot classification and retrieval tasks.