Few-shot Learning Methods Research Articles

Explicable Fine-Grained Aircraft Recognition Via Deep Part Parsing Prior Framework for High-Resolution Remote Sensing Imagery.

Aircraft recognition is crucial in both civil and military fields, and high-spatial resolution remote sensing has emerged as a practical approach. However, existing data-driven methods fail to locate discriminative regions for effective feature extraction due to limited training data, leading to poor recognition performance. To address this issue, we propose a knowledge-driven deep learning method called the explicable aircraft recognition framework based on a part parsing prior (APPEAR). APPEAR explicitly models the aircraft's rigid structure as a pixel-level part parsing prior, dividing it into five parts: 1) the nose; 2) left wing; 3) right wing; 4) fuselage; and 5) tail. This fine-grained prior provides reliable part locations to delineate aircraft architecture and imposes spatial constraints among the parts, effectively reducing the search space for model optimization and identifying subtle interclass differences. A knowledge-driven aircraft part attention (KAPA) module uses this prior to achieving a geometric-invariant representation for identifying discriminative features. Part features are generated by part indexing in a specific order and sequentially embedded into a compact space to obtain a fixed-length representation for each part, invariant to aircraft orientation and scale. The part attention module then takes the embedded part features, adaptively reweights their importance to identify discriminative parts, and aggregates them for recognition. The proposed APPEAR framework is evaluated on two aircraft recognition datasets and achieves superior performance. Moreover, experiments with few-shot learning methods demonstrate the robustness of our framework in different tasks. Ablation analysis illustrates that the fuselage and wings of the aircraft are the most effective parts for recognition.

Feature Mixture on Pre-Trained Model for Few-Shot Learning.

Few-shot learning (FSL) aims at recognizing a novel object under limited training samples. A robust feature extractor (backbone) can significantly improve the recognition performance of the FSL model. However, training an effective backbone is a challenging issue since 1) designing and validating structures of backbones are time-consuming and expensive processes, and 2) a backbone trained on the known (base) categories is more inclined to focus on the textures of the objects it learns, which is hard to describe the novel samples. To solve these problems, we propose a feature mixture operation on the pre-trained (fixed) features: 1) We replace a part of the values of the feature map from a novel category with the content of other feature maps to increase the generalizability and diversity of training samples, which avoids retraining a complex backbone with high computational costs. 2) We use the similarities between the features to constrain the mixture operation, which helps the classifier focus on the representations of the novel object where these representations are hidden in the features from the pre-trained backbone with biased training. Experimental studies on five benchmark datasets in both inductive and transductive settings demonstrate the effectiveness of our feature mixture (FM). Specifically, compared with the baseline on the Mini-ImageNet dataset, it achieves 3.8% and 4.2% accuracy improvements for 1 and 5 training samples, respectively. Additionally, the proposed mixture operation can be used to improve other existing FSL methods based on backbone training.

Prototypical Network with Residual Attention for Modulation Classification of Wireless Communication Signals

Automatic modulation classification (AMC) based on data-driven deep learning (DL) can achieve excellent classification performance. However, in the field of electronic countermeasures, it is difficult to extract salient features from wireless communication signals under scarce samples. Aiming at the problem of modulation classification under scarce samples, this paper proposes a few-shot learning method using prototypical network (PN) with residual attention (RA), namely PNRA, to achieve the AMC. Firstly, the RA is utilized to extract the feature vector of wireless communication signals. Subsequently, the feature vector is mapped to a new feature space. Finally, the PN is utilized to measure the Euclidean distance between the feature vector of the query point and each prototype in this space, determining the type of the signals. In comparison to mainstream few-shot learning (FSL) methods, the proposed PNRA can achieve effective and robust AMC under the data-hungry condition.

Relational concept enhanced prototypical network for incremental few-shot relation classification

Compared with conventional close-domain relation classification, incremental few-shot relation classification requires incrementally to learn novel relations through very few samples without forgetting base relations, which is more fitting to the needs of real scenarios. Existing incremental few-shot learning methods commonly leverage memory to freeze knowledge of the learned base classes, and then incrementally learn new classes from very few labeled samples. However, there are still two challenges to these existing approaches: overfitting novel relations and the catastrophic forgetting of base relations. In this paper, we propose a relational concept enhanced prototypical network (RC-EProto) to address these problems. First, we build a concept alignment module to eliminate the difference between relational concepts and samples, and we incorporate relational concepts for sample representation by this module to construct a more appropriate prototype. Second, we develop a contrastive approach on the base relations to address the forgetting of old knowledge. The samples of the base relation are pulled tightly with its prototype in the embedding space, while the samples of the other relations are pulled far from this base prototype. Extensive experiments on FewRel 1.0 and 2.0 datasets demonstrate the superiority of our method compared with the state-of-the-art methods. Notably, RC-EProto maintains an excellent incremental few-shot domain adaptation learning ability. Compared to existing methods, our model is able to achieve near-optimal results for the base relations while improving the classification accuracy of the novel relations by 13.48% and 9.63% for the 1-increment and 5-increment settings on FewRel 2.0, respectively.

LibFewShot: A Comprehensive Library for Few-Shot Learning.

Few-shot learning, especially few-shot image classification, has received increasing attention and witnessed significant advances in recent years. Some recent studies implicitly show that many generic techniques or "tricks", such as data augmentation, pre-training, knowledge distillation, and self-supervision, may greatly boost the performance of a few-shot learning method. Moreover, different works may employ different software platforms, backbone architectures and input image sizes, making fair comparisons difficult and practitioners struggle with reproducibility. To address these situations, we propose a comprehensive library for few-shot learning (LibFewShot) by re-implementing eighteen state-of-the-art few-shot learning methods in a unified framework with the same single codebase in PyTorch. Furthermore, based on LibFewShot, we provide comprehensive evaluations on multiple benchmarks with various backbone architectures to evaluate common pitfalls and effects of different training tricks. In addition, with respect to the recent doubts on the necessity of meta- or episodic-training mechanism, our evaluation results confirm that such a mechanism is still necessary especially when combined with pre-training. We hope our work can not only lower the barriers for beginners to enter the area of few-shot learning but also elucidate the effects of nontrivial tricks to facilitate intrinsic research on few-shot learning. The source code is available from https://github.com/RL-VIG/LibFewShot.

Dynamic feature splicing for few-shot rare disease diagnosis.

Annotated images for rare disease diagnosis are extremely hard to collect. Therefore, identifying rare diseases under a few-shot learning (FSL) setting is significant. Existing FSL methods transfer useful and global knowledge from base classes with abundant training samples to enrich features of novel classes with few training samples, but still face difficulties when being applied to medical images due to the complex lesion characteristics and large intra-class variance. In this paper, we propose a dynamic feature splicing (DNFS) framework for few-shot rare disease diagnosis. Under DNFS, both low-level features (i.e., the output of three convolutional blocks) and high-level features (i.e., the output of the last fully connected layer) of novel classes are dynamically enriched. We construct the position coherent DNFS (P-DNFS) module to perform low-level feature splicing, where a lesion-oriented Transformer is designed to detect lesion regions. Thus, novel-class channels are replaced by similar base-class channels within the detected lesion regions to achieve disease-related feature enrichment. We also devise a semantic coherent DNFS (S-DNFS) module to perform high-level feature splicing. It explores cross-image channel relations and selects base-class channels with semantic consistency for explicit knowledge transfer. Both low-level and high-level feature splicings are performed dynamically and iteratively. Consequently, abundant spliced features are generated for disease diagnosis, leading to more accurate decision boundary and improved diagnosis performance. Extensive experiments have been conducted on three medical image classification datasets. Our results suggest that the proposed DNFS achieves superior performance against state-of-the-art approaches.

A Large-Class Few-Shot Learning Method Based on High-Dimensional Features

Large-class few-shot learning has a wide range of applications in many fields, such as the medical, power, security, and remote sensing fields. At present, many few-shot learning methods for fewer-class scenarios have been proposed, but little research has been performed for large-class scenarios. In this paper, we propose a large-class few-shot learning method called HF-FSL, which is based on high-dimensional features. Recent theoretical research shows that if the distribution of samples in a high-dimensional feature space meets the conditions of compactness within the class and the dispersion between classes, the large-class few-shot learning method has a better generalization ability. Inspired by this theory, the basic idea is use a deep neural network to extract high-dimensional features and unitize them to project the samples onto a hypersphere. The global orthogonal regularization strategy can then be used to make samples of different classes on the hypersphere that are as orthogonal as possible, so as to achieve the goal of sample compactness within the class and the dispersion between classes in high-dimensional feature space. Experiments on Omniglot, Fungi, and ImageNet demonstrate that the proposed method can effectively improve the recognition accuracy in a large-class FSL problem.

From patch, sample to domain: Capture geometric structures for few-shot learning

Few-shot learning aims to recognize novel concepts with only few samples by using prior knowledge learned from the seen concepts. In this paper, we address the problem of few-shot learning under domain shifts. Traditional few-shot learning methods are not directly applicable to cross-domain scenarios due to the large discrepancy of feature distributions across domains. To this end, we propose a novel Hierarchical Optimal Transport network with Attention (HOTA) for cross-domain few-shot learning. The underlying idea is to learn the transferable and discriminative embeddings by taking advantage of the hierarchical geometric structures among image data, ranging from patch, sample to domain. The HOTA framework utilizes a hierarchical optimal transport network to smooth the domain shifts by domain alignment while enhancing the discrimination and the transferability of the embeddings by aligning the patches of images. To further enhance the transferability, HOTA conducts a mix-up data augmentation based on cross-domain attention to capture the relationships of samples in different domains. The extensive experiments on a variety of few-shot benchmark scenarios demonstrate that HOTA outperforms the state-of-the-art methods under both supervised and unsupervised conditions.

Transformers, convolutional neural networks, and few-shot learning for classification of histopathological images of oral cancer

The diagnosis of oral squamous cell carcinoma or oral leukoplakia and the presence or absence of oral epithelial dysplasia is carried by pathologists. In recent years, deep learning has been presented to deal with the automated detection of various pathologies using digital images. One of the main limitations to applying deep learning to histopathological images is the lack of public datasets. In order to fill this gap, a joint effort has been made and a new dataset of histopathological images of oral cancer, named P-NDB-UFES, has been collected, annotated, and analyzed by oral pathologists generating the gold-standard for classification. This dataset is composed of 3763 images of patches of histopathological images with oral squamous cell carcinoma (29%), dysplasia (51.29%), and without dysplasia (18.79%). Next, convolutional neural network (CNN), transformers neural networks, and few-shot learning approaches (i.e., Siamese, Triplet, and ProtoNet) were investigated to classify oral squamous cell carcinoma and the presence or absence of oral dysplasia. Experimental results indicate that the CNNs and transformers models, in general, have no statistically significant difference, with only DenseNet-121 outperforming transformers at a balanced accuracy (BCC) of 91.91%, recall, and precision of 91.93%. Few-shot learning methods were inferior when compared to other methods, with different configurations having statistical differences among themselves. For ProtoNet architectures, the usage of hyperbolic space showed to have a similar behavior to Euclidean distance, however, these results were heavily influenced by the optimizer used.

CPCL: Conceptual prototypical contrastive learning for Few-Shot text classification

As prototype-based Few-Shot Learning methods, Prototypical Network generates prototypes for each class in a low-resource state and classify by a metric module. Therefore, the quality of prototypes matters but they are inaccurate from the few support instances, and the domain-specific information of training data are harmful to the generalizability of prototypes. We propose a Conceptual Prototype (CP), which contains both rich instance and concept features. The numerous query data can inspire the few support instances. An interactive network is designed to leverage the interrelation between support set and query-detached set to acquire a rich Instance Prototype which is typical on the whole data. Besides, class labels are introduced to prototype by prompt engineering, which makes it more conceptual. The label-only concept makes prototype immune to domain-specific information in training phase to improve its generalizability. Based on CP, Conceptual Prototypical Contrastive Learning (CPCL) is proposed where PCL brings instances closer to its corresponding prototype and pushes away from other prototypes. “2-way 5-shot” experiments show that CPCL achieves 92.41% accuracy on ARSC dataset, 2.30% higher than other prototype-based models. Meanwhile, the 0-shot performance of CPCL is comparable to Induction Network in the 5-shot way, indicating that our model is adequate for 0-shot tasks.

Few-shot learning for joint model in underwater acoustic target recognition

In underwater acoustic target recognition, there is a lack of massive high-quality labeled samples to train robust deep neural networks, and it is difficult to collect and annotate a large amount of base class data in advance unlike the image recognition field. Therefore, conventional few-shot learning methods are difficult to apply in underwater acoustic target recognition. In this report, following advanced self-supervised learning frameworks, a learning framework for underwater acoustic target recognition model with few samples is proposed. Meanwhile, a semi-supervised fine-tuning method is proposed to improve the fine-tuning performance by mining and labeling partial unlabeled samples based on the similarity of deep features. A set of small sample datasets with different amounts of labeled data are constructed, and the performance baselines of four underwater acoustic target recognition models are established based on these datasets. Compared with the baselines, using the proposed framework effectively improves the recognition effect of four models. Especially for the joint model, the recognition accuracy has increased by 2.04% to 12.14% compared with the baselines. The model performance on only 10 percent of the labeled data can exceed that on the full dataset, effectively reducing the dependence of model on the number of labeled samples. The problem of lack of labeled samples in underwater acoustic target recognition is alleviated.

Sparse spatial transformers for few-shot learning

Learning from limited data is challenging because data scarcity leads to a poor generalization of the trained model. A classical global pooled representation will probably lose useful local information. Many few-shot learning methods have recently addressed this challenge using deep descriptors and learning a pixel-level metric. However, using deep descriptors as feature representations may lose image contextual information. Moreover, most of these methods independently address each class in the support set, which cannot sufficiently use discriminative information and task-specific embeddings. In this paper, we propose a novel transformer-based neural network architecture called sparse spatial transformers (SSFormers), which finds task-relevant features and suppresses task-irrelevant features. Particularly, we first divide each input image into several image patches of different sizes to obtain dense local features. These features retain contextual information while expressing local information. Then, a sparse spatial transformer layer is proposed to find spatial correspondence between the query image and the full support set to select task-relevant image patches and suppress task-irrelevant image patches. Finally, we propose using an image patch-matching module to calculate the distance between dense local representations, thus determining which category the query image belongs to in the support set. Extensive experiments on popular few-shot learning benchmarks demonstrate the superiority of our method over state-of-the-art methods. Our source code is available at \url{https://github.com/chenhaoxing/ssformers}.

Hyperplane projection network for few-shot relation classification

Recently meta-learning-based few-shot learning methods have been widely used for relation classification. Previous work reveals that meta-learning performs poorly in scenarios where the edge probability distribution of the target domain dataset appears to be significantly different from the source domain. In this paper, we enhance the meta-learning framework with high-dimensional semantic feature extraction and hyperplane projection metrics for meta-tasks. First, we enhance the focus of BERT on entity words by adding entity markers and vector pooling. After that, the high-dimensional semantic features of the support set are extracted and transformed into hyperplanes. Finally, we obtain the classification results by calculating the projection distance between the query sample and the hyperplane. In addition, we design a auxiliary function with a plane correction factor, which can better amplify the plane spacing and reduce the degree of category confusion, which is important for solving the problem of metric spatial loss. Experiments on two real-world few-shot datasets show that our model HPN is more effective in classifying few-shot relations in the same domain and domain-adapted scenarios. And HPN is more stable on NOTA tasks.

Prototype Completion for Few-Shot Learning.

Few-shot learning (FSL) aims to recognize novel classes with few examples. Pre-training based methods effectively tackle the problem by pre-training a feature extractor and then fine-tuning it through the nearest centroid based meta-learning. However, results show that the fine-tuning step makes marginal improvements. In this paper, 1) we figure out the reason, i.e., in the pre-trained feature space, the base classes already form compact clusters while novel classes spread as groups with large variances, which implies that fine-tuning feature extractor is less meaningful; 2) instead of fine-tuning feature extractor, we focus on estimating more representative prototypes. Consequently, we propose a novel prototype completion based meta-learning framework. This framework first introduces primitive knowledge (i.e., class-level part or attribute annotations) and extracts representative features for seen attributes as priors. Second, a part/attribute transfer network is designed to learn to infer the representative features for unseen attributes as supplementary priors. Finally, a prototype completion network is devised to learn to complete prototypes with these priors. Moreover, to avoid the prototype completion error, we further develop a Gaussian based prototype fusion strategy that fuses the mean-based and completed prototypes by exploiting the unlabeled samples. At last, we also develop an economic prototype completion version for FSL, which does not need to collect primitive knowledge, for a fair comparison with existing FSL methods without external knowledge. Extensive experiments show that our method: (i) obtains more accurate prototypes; (ii) achieves superior performance on both inductive and transductive FSL settings. Our codes are open-sourced at https://github.com/zhangbq-research/Prototype_Completion_for_FSL.

Hierarchical Knowledge Propagation and Distillation for Few-Shot Learning

Recent research efforts on Few-Shot Learning (FSL) have achieved extensive progress. However, the existing efforts primarily focus on the transductive setting of FSL, which is heavily challenged by the limited quantity of the unlabeled query set. Although a few inductive-based FSL methods have been studied, most of them emphasize learning superb feature extraction networks. As a result, they may ignore the relations between sample-level and class-level representations, which are particularly crucial when labeled samples are scarce. This paper proposes an inductive FSL framework that leverages the Hierarchical Knowledge Propagation and Distillation, named HKPD. To learn more discriminative sample-level representations, HKPD first constructs a sample-level information propagation module that explores pairwise sample relations. Subsequently, a class-level information propagation module is designed to obtain and update the class-level information. Moreover, a self-distillation module is adopted to further improve the learned representations by propagating the obtained knowledge across this hierarchical architecture. Extensive experiments conducted on the commonly used few-shot benchmark datasets demonstrate the superiority of the proposed HKPD method, which outperforms the current state-of-the-art methods.

Improving Plant Disease Recognition Through Gradient-Based Few-shot Learning with Attention Mechanisms

This study investigates the use of few-shot learning algorithms to improve classification performance in situations where traditional deep learning methods fail due to a lack of training data. Specifically, we propose a few-shot learning approach using the Almost No Inner Loop (ANIL) algorithm and attention modules to classify tomato diseases in the Plant Village dataset. The attended features obtained from the five separate attention modules are classified using a Multi Layer Perceptron (MLP) classifier, and the soft voting method is used to weigh the classification scores from each classifier. The results demonstrate that our proposed approach achieves state-of-the-art accuracy rates of 97.05% and 97.66% for 10-shot and 20-shot classification, respectively. Our approach demonstrates the potential for incorporating attention mechanisms in feature extraction processes and suggests new avenues for research in few-shot learning methods.

A lightweight dense relation network with attention for hyperspectral image few-shot classification

Deep learning methods have significantly progressed in hyperspectral image (HSI) classification. However, deep learning relies on large labeled data for training. The cost of labeling samples is enormous. In practical classification tasks, only a few labeled samples are available. A lightweight dense relation network with attention (LDA-RN) was presented for HSI few-shot classification. First, we design a 3D feature embedding module for spectral–spatial feature extraction. A light attention module (CSA), including the channel attention (CA) submodule and spatial attention (SA) submodule, is applied to reinforce essential features. Dense connections are used to fully utilize the feature information in each layer. Second, we propose a 2D relation learning module for classification by comparing the similarity between samples, using global average pooling to reduce the model parameters. Finally, meta-learning and fine-tuning training techniques are employed to enhance the model’s generalization ability and reduce the dependence on the labeled samples. Experimental results on three available HSI datasets suggest that the proposed LDA-RN outperforms existing advanced few-shot learning methods. For example, the experimental results show that the overall accuracy of the proposed method improves by 5.67% on the Univerisity of Pavia data set with only five labeled samples compared to the best method DCFSL.

Task‐oriented feature hallucination for few‐shot image classification

AbstractData hallucination generates additional training examples for novel classes to alleviate the data scarcity problem in few‐shot learning (FSL). Existing hallucination‐based FSL methods normally train a general embedding model first by applying information extracted from base classes that have abundant data. In those methods, hallucinators are then built upon the trained embedding model to generate data for novel classes. However, these hallucination methods usually rely on general‐purpose embeddings, limiting their ability to generate task‐oriented samples for novel classes. Recent studies have shown that task‐specific embedding models, which are adapted to novel tasks, can achieve better classification performance. To improve the performance of example hallucination for tasks, a task‐oriented embedding model is used in the proposed method to perform task‐oriented generation. After the initialization, the hallucinator is finetuned by applying a task‐oriented embedding model with the guidance of a teacher–student mechanism. The proposed task‐oriented hallucination method contains two steps. An initial embedding network and an initial hallucinator are trained with a base dataset in the first step. The second step contains a pseudo‐labelling process where the base dataset is pseudo‐labelled using support data of the few‐shot task and a task‐oriented fine‐tuning process where the embedding network and hallucinator are adjusted simultaneously. Both the embedding network and the hallucinator are updated with the support set and the pseudo‐labelled base dataset using knowledge distillation. The experiments are conducted on four popular few‐shot datasets. The results demonstrate that the proposed approach outperforms state‐of‐the‐art methods with 0.8% to 4.08% increases in classification accuracy for 5‐way 5‐shot tasks. It also achieves comparable accuracy to state‐of‐the‐art methods for 5‐way 1‐shot tasks.

SpatialFormer: Semantic and Target Aware Attentions for Few-Shot Learning

Recent Few-Shot Learning (FSL) methods put emphasis on generating a discriminative embedding features to precisely measure the similarity between support and query sets. Current CNN-based cross-attention approaches generate discriminative representations via enhancing the mutually semantic similar regions of support and query pairs. However, it suffers from two problems: CNN structure produces inaccurate attention map based on local features, and mutually similar backgrounds cause distraction. To alleviate these problems, we design a novel SpatialFormer structure to generate more accurate attention regions based on global features. Different from the traditional Transformer modeling intrinsic instance-level similarity which causes accuracy degradation in FSL, our SpatialFormer explores the semantic-level similarity between pair inputs to boost the performance. Then we derive two specific attention modules, named SpatialFormer Semantic Attention (SFSA) and SpatialFormer Target Attention (SFTA), to enhance the target object regions while reduce the background distraction. Particularly, SFSA highlights the regions with same semantic information between pair features, and SFTA finds potential foreground object regions of novel feature that are similar to base categories. Extensive experiments show that our methods are effective and achieve new state-of-the-art results on few-shot classification benchmarks.

Bi-directional Feature Reconstruction Network for Fine-Grained Few-Shot Image Classification

The main challenge for fine-grained few-shot image classification is to learn feature representations with higher inter-class and lower intra-class variations, with a mere few labelled samples. Conventional few-shot learning methods however cannot be naively adopted for this fine-grained setting -- a quick pilot study reveals that they in fact push for the opposite (i.e., lower inter-class variations and higher intra-class variations). To alleviate this problem, prior works predominately use a support set to reconstruct the query image and then utilize metric learning to determine its category. Upon careful inspection, we further reveal that such unidirectional reconstruction methods only help to increase inter-class variations and are not effective in tackling intra-class variations. In this paper, we for the first time introduce a bi-reconstruction mechanism that can simultaneously accommodate for inter-class and intra-class variations. In addition to using the support set to reconstruct the query set for increasing inter-class variations, we further use the query set to reconstruct the support set for reducing intra-class variations. This design effectively helps the model to explore more subtle and discriminative features which is key for the fine-grained problem in hand. Furthermore, we also construct a self-reconstruction module to work alongside the bi-directional module to make the features even more discriminative. Experimental results on three widely used fine-grained image classification datasets consistently show considerable improvements compared with other methods. Codes are available at: https://github.com/PRIS-CV/Bi-FRN.