Few-shot learning (FSL) is a challenging yet promising technique that aims to discriminate objects based on a few labeled examples. Learning a high-quality feature representation is key with few-shot data, and many existing models attempt to extract general information from the sample or task levels. However, the common sample-level means of feature representation limits the models generalizability to different tasks, while task-level representation may lose class characteristics due to excessive information aggregation. In this article, we synchronize the class-specific and task-shared information from the class and task levels to obtain a better representation. Structure-based contrastive learning is introduced to obtain class-specific representations by increasing the interclass distance. A hierarchical class structure is constructed by clustering semantically similar classes using the idea of granular computing. When guided by a class structure, it is more difficult to distinguish samples in different classes that have similar characteristics than those with large interclass differences. To this end, structure-guided contrastive learning is introduced to study class-specific information. A hierarchical graph neural network is established to transfer task-shared information from coarse to fine. It hierarchically infers the target sample based on all samples in the task and yields a more general representation for FSL classification. Experiments on four benchmark datasets demonstrate the advantages of our model over several state-of-the-art models.

The significant success of machine learning models is mainly based on a large amount of data for training iterations, but this limits their generalization for few-shot data. Some existing models utilize the extensive visual and textual modal knowledge of vision-language pre-trained models (VLPs) to compensate for the data scarcity problem. However, they may suffer from a classification bias problem during the fusion of multi-modal information since that they focus on the inter-modal matching while neglecting intra-modal recognition for few-shot images. In this paper, we propose a novel few-shot model with mixed-modal prototypes by partial-tuning the VLPs for better information fusion. It aims to yield a high-quality class prototype representation by integrating the abundant multi-modal knowledge of VLPs and the specific-task information of low-shot visual data. Specifically, we introduce an image-text alignment module to ensure the consistency of the few-shot visual representation and the textual knowledge of VLPs at the feature space. A self-similar learning module is designed to excavate the local and detailed characters of specific class, which is crucial under the data scarcity. Additionally, to preserve the generalizable pre-trained knowledge in the maximum extent, we partial-tune the parameters of VLPs to adapt for the few-shot tasks. To sum up, we mix multi-modal information at the feature representation level instead of fusing multi-modal matching similarities, which effectively mitigates classification bias and ultimately enhances the model performance for few-shot data. The extensive experiments are conducted to evaluate the effectiveness of our model on 11 benchmark datasets and the results show its promising.

With the availability of limited data, mislabeled data makes few-shot noise learning more complex and challenging. However, existing models predominantly focus on mitigating noise solely during testing, neglecting the noise contamination during model training, which directly introduces more uncertainty and ambiguity into the model especially in low-shot training. In this paper, we propose a novel few-shot framework via the dual-branch adversarial noise pretraining strategy (FS-DANP) to combat label noise. FS-DANP aims to comprehensively mitigate the noise interference during both training and testing pipelines, enhancing model performance and reliability. Initially, we introduce a dual-branch adversarial noise pretraining strategy to alleviate the uncertainty and ambiguity, while ensuring a discriminative representation learning with limited data against the training noise. Additionally, a weighted prototype few-shot classifier is designed to remove noise guided by the pretrained model, enabling that the classifier can quickly learn new concepts from limited labeled examples. Extensive experimental evaluations validate the superiority and robustness of our model in navigating diverse noise environments, demonstrating its exceptional performance and adaptability.

Large-scale vision-language pre-trained models like CLIP are extensively employed in few-shot tasks due to their robust generalization capabilities. Existing methods usually incorporate additional techniques to acquire knowledge for new tasks building upon the general knowledge in CLIP. However, they do not realize that the task-related knowledge might be implicitly embedded within the general knowledge well-learned. In this paper, we propose a novel framework to reallocate and evolve the general knowledge for specific few-shot tasks (REGK), mimicking the human "Attention Allocation" cognition mechanism. With a learnable mask-tuning selection, REGK focuses on selecting the task-related parameters of CLIP while learning specific few-shot knowledge without altering CLIP underlying framework. Specifically, we initially observe that inheriting the strong knowledge representation capability in CLIP is more advantageous for few-shot learning than its task-solving ability. Subsequently, a two-stage tuning framework is introduced to reallocate and control the mask-tuning on different tasks. It allows model automatically mask-tuning on different few-shot tasks with selective sparsity training. In this way, we achieve reliable transfer of task-related knowledge and effective exploration of new knowledge from limited data to enhance few-shot learning. Extensive experiments validate the superiority and potentiality of our model.

Few-shot learning (FSL) is a challenging task in classifying new classes from few labelled examples. Many existing models embed class structural knowledge as prior knowledge to enhance FSL against data scarcity. However, they fall short of connecting the class structural knowledge with the limited visual information which plays a decisive role in FSL model performance. In this paper, we propose a unified FSL framework with multi-granularity knowledge fusion and decision-making (MGKFD) to overcome the limitation. We aim to simultaneously explore the visual information and structural knowledge, working in a mutual way to enhance FSL. On the one hand, we strongly connect global and local visual information with multigranularity class knowledge to explore intra-image and inter-class relationships, generating specific multi-granularity class representations with limited images. On the other hand, a weight fusion strategy is introduced to integrate multi-granularity knowledge and visual information to make the classification decision of FSL. It enables models to learn more effectively from limited labelled examples and allows generalization to new classes. Moreover, considering varying erroneous predictions, a hierarchical loss is established by structural knowledge to minimize the classification loss, where greater degree of misclassification is penalized more. Experimental results on three benchmark datasets show the advantages of MGKFD over several advanced models.

Few-shot learning, which aims to identify new classes with very few samples, is an increasingly popular and crucial research topic in the machine learning. Many models use distance measurement to determine similarities among single samples and achieve accurate classification results. However, distance calculations incur substantial costs and time based on a single sample, and the linear measurement model cannot accurately represent the differences and connections between samples. This paper proposes a coarse-grained granulation relation network (CGRN) model for few-shot classification. First, all the single samples of each class are clustered into coarse grain to represent the feature information of all the class samples, which can significantly reduce computational complexities. Second, a relation network is built to measure the degree of similarity among the test samples and the coarse grain obtained above, which can reveal the differences and connections between the samples. The experimental results demonstrate that this model outperforms some popular distance measurement-based few-shot learning models. For example, CGRN is at least 0.5% better than other models in 20-way 5-shot on the Omniglot dataset and achieves 0.8% improvement over the second-best model in 5-way 1-shot on the tiered-ImageNet dataset.

Few-shot learning aims to recognize new visual concepts with a small number of labeled samples. The hierarchical structure based on inter-class labels performs well in many few-shot learning models. However, intra-class features are similar and difficult to distinguish, which is important for mining the correlation and independence between intra-class features in the scene of sparse data. In this paper, we propose a few-shot learning model with a self-similarity feature representation by a hierarchical relation network, which considers inter-class labels and intra-class features to guide few-shot learning. First, we introduce a self-similarity feature representation module as the intermediate feature transform in the neural network. Unlike the traditional model, it extracts specific feature information from intra-class features. Second, we leverage the inter-class label hierarchical structure as important auxiliary information to establish a hierarchical relation network metric module. The module uses coarse-grained information to guide fine-grained classification, which effectively alleviates the problem of insufficient data. Experimental results show that our model improves the classification accuracy, reaching 58.68% on the tieredImageNet dataset.

Few-shot learning is one of the significant areas of machine learning, which aims to recognize novel visual classes from few labeled examples. Many existing models make full use of the similarity of inner-class features and achieve satisfactory results. However, these models assume that classes are independent of each other, ignoring the inter-class relationship. In this paper, we propose a few-shot hierarchical classification model via multi-granularity relation networks (HMRN) considering both the inner-class similarity and inter-class relationship. The multi-granularity relationship among coarse- and fine-grained classes is an important auxiliary information in the class hierarchical structure originated from data. Thus, we first extract hierarchical features of different granularity classes according to the membership relationship among the classes. Second, we build multi-granularity relation networks to obtain the inner-class similarity relation of different granularity classes using the hierarchical features. Finally, we consider the tradeoff among the inner-class similarity relation of different granularity classes for hierarchical few-shot learning, which takes the information of coarse-grained classes to assist the learning of fine-grained classes. Experimental results show that our model outperforms several state-of-the-art flat (without hierarchical structure) models and hierarchical models. For example, the accuracy of HMRN is about 3.00% better than that of flat models on the tieredImageNet dataset.

Long-tailed distribution learning is one of the critical research fields of deep learning and has gradually become a research hotspot. Existing re-sampling methods for long-tailed data classification attempt to adjust the number of tail class samples to balance the overall feature space and achieve satisfactory results. However, the methods impair the representative ability of the learned features to a certain extent, which in turn affects the tail class feature space. In this paper, we propose a hybrid ResNet based on joint basic and attention modules to enhance the tail class feature space, which provides rich discriminative and representative features in the tail class feature space. Firstly, we use hybrid ResNet to extract features, where the basic module ResNet and the attention module ResNet extract head and tail class features, respectively. The enhancement of tail class features can reduce the dependence of the classifier on head class features. Secondly, we build a fusion loss function, which considers the tradeoff between head loss and tail loss for long-tailed distribution learning. Experimental results show that the proposed model outperforms several state-of-the-art models in the long-tailed classification. Our model was 2.67% better than the optimal method under the long-tailed Tiny-Imagenet-LT dataset with an imbalanced ratio of 100.