Filter Bank Networks for Few-Shot Class-Incremental Learning

Deep Convolution Neural Networks(DCNNs)can capture discriminative features from large datasets.However,how to incrementally learn new samples without forgetting old ones and recognize novel classes that arise in the dynamically changing world,e.g.,classifying newly discovered fish species,remains an open problem.We address an even more challenging and realistic setting of this problem where new class samples are insufficient,i.e.,Few-Shot Class-Incremental Learning(FSCIL).Current FSCIL methods augment the training data to alleviate the overfitting of novel classes.By contrast,we propose Filter Bank Networks(FBNs)that augment the learnable filters to capture fine-detailed features for adapting to future new classes.In the forward pass,FBNs augment each convolutional filter to a virtual filter bank containing the canonical one,i.e.,itself,and multiple transformed versions.During back-propagation,FBNs explicitly stimulate fine-detailed features to emerge and collectively align all gradients of each filter bank to learn the canonical one.FBNs capture pattern variants that do not yet exist in the pretraining session,thus making it easy to incorporate new classes in the incremental learning phase.Moreover,FBNs introduce model-level prior knowledge to efficiently utilize the limited few-shot data.Extensive experiments on MNIST,CIFAR100,CUB200,andMini-ImageNet datasets show that FBNs consistently outperformthe baseline by a significantmargin,reporting new state-of-the-art FSCIL results.In addition,we contribute a challenging FSCIL benchmark,Fishshot1K,which contains 8261 underwater images covering 1000 ocean fish species.The code is included in the supplementary materials.

Squeezing More Past Knowledge for Online Class-Incremental Continual Learning

Continual learning(CL)studies the problem of learning to accumulate knowledge over time from a stream of data.A crucial challenge is that neural networks suffer from performance degradation on previously seen data,known as catastrophic forgetting,due to allowing parameter sharing.In this work,we consider a more practical online class-incremental CL setting,where the model learns new samples in an online manner and may continuously experience new classes.Moreover,prior knowledge is unavailable during training and evaluation.Existing works usually explore sample usages from a single dimension,which ignores a lot of valuable supervisory information.To better tackle the setting,we propose a novel replay-based CL method,which leverages multi-level representations produced by the intermediate process of training samples for replay and strengthens supervision to consolidate previous knowledge.Specifically,besides the previous raw samples,we store the corresponding logits and features in the memory.Furthermore,to imitate the prediction of the past model,we construct extra constraints by leveraging multi-level information stored in the memory.With the same number of samples for replay,our method can use more past knowledge to prevent interference.We conduct extensive evaluations on several popular CL datasets,and experiments show that our method consistently outperforms state-of-the-art methods with various sizes of episodic memory.We further provide a detailed analysis of these results and demonstrate that our method is more viable in practical scenarios.

Few-shot working condition recognition of a sucker-rod pumping system based on a 4-dimensional time-frequency signature and meta-learning convolutional shrinkage neural network

The accurate and intelligent identification of the working conditions of a sucker-rod pumping system is necessary. As onshore oil extraction gradually enters its mid-to late-stage, the cost required to train a deep learning working condition recognition model for pumping wells by obtaining enough new working condition samples is expensive. For the few-shot problem and large calculation issues of new working conditions of oil wells, a working condition recognition method for pumping unit wells based on a 4-dimensional time-frequency signature (4D-TFS) and meta-learning convolutional shrinkage neural network (ML-CSNN) is proposed. First, the measured pumping unit well workup data are converted into 4D-TFS data, and the initial feature extraction task is performed while compressing the data. Subsequently, a convolutional shrinkage neural network (CSNN) with a specific structure that can ablate low-frequency features is designed to extract working conditions features. Finally, a meta-learning fine-tuning framework for learning the network parameters that are susceptible to task changes is merged into the CSNN to solve the few-shot issue. The results of the experiments demonstrate that the trained ML-CSNN has good recognition accuracy and generalization ability for few-shot working condition recognition. More specifically, in the case of lower computational complexity, only few-shot samples are needed to fine-tune the network parameters, and the model can be quickly adapted to new classes of well conditions.

Automated Classification of Inherited Retinal Diseases in Optical Coherence Tomography Images Using Few-shot Learning

Objective To develop a few-shot learning(FSL) approach for classifying optical coherence tomography(OCT) images in patients with inherited retinal disorders(IRDs).Methods In this study, an FSL model based on a student–teacher learning framework was designed to classify images. 2,317 images from 189 participants were included. Of these, 1,126 images revealed IRDs, 533 were normal samples, and 658 were control samples.Results The FSL model achieved a total accuracy of 0.974–0.983, total sensitivity of 0.934–0.957, total specificity of 0.984–0.990, and total F1 score of 0.935–0.957, which were superior to the total accuracy of the baseline model of 0.943–0.954, total sensitivity of 0.866–0.886, total specificity of 0.962–0.971,and total F1 score of 0.859–0.885. The performance of most subclassifications also exhibited advantages. Moreover, the FSL model had a higher area under curves(AUC) of the receiver operating characteristic(ROC) curves in most subclassifications.Conclusion This study demonstrates the effective use of the FSL model for the classification of OCT images from patients with IRDs, normal, and control participants with a smaller volume of data. The general principle and similar network architectures can also be applied to other retinal diseases with a low prevalence.

SW-Net: A novel few-shot learning approach for disease subtype prediction

Few-shot learning is becoming more and more popular in many fields,especially in the computer vision field.This inspires us to introduce few-shot learning to the genomic field,which faces a typical few-shot problem because some tasks only have a limited number of samples with high-dimensions.The goal of this study was to investigate the few-shot disease sub-type prediction problem and identify patient subgroups through training on small data.Accurate disease subtype classification allows clinicians to efficiently deliver investigations and interventions in clinical practice.We propose the SW-Net,which simulates the clinical process of extracting the shared knowledge from a range of interrelated tasks and generalizes it to unseen data.Our model is built upon a simple baseline,and we modified it for genomic data.Supportbased initialization for the classifier and transductive fine-tuning techniques were applied in our model to improve prediction accuracy,and an Entropy regularization term on the query set was appended to reduce over-fitting.Moreover,to address the high dimension and high noise issue,we future extended a feature selection module to adaptively select important features and a sample weighting module to prioritize high-confidence samples.Experiments on simulated data and The Cancer Genome Atlas meta-dataset show that our new baseline model gets higher prediction accuracy compared to other competing algorithms.

Dynamic Analogical Association Algorithm Based on Manifold Matching for Few-Shot Learning

At present,deep learning has been well applied in many fields.However,due to the high complexity of hypothesis space,numerous training samples are usually required to ensure the reliability of minimizing experience risk.Therefore,training a classifier with a small number of training examples is a challenging task.From a biological point of view,based on the assumption that rich prior knowledge and analogical association should enable human beings to quickly distinguish novel things from a few or even one example,we proposed a dynamic analogical association algorithm to make the model use only a few labeled samples for classification.To be specific,the algorithm search for knowledge structures similar to existing tasks in prior knowledge based on manifold matching,and combine sampling distributions to generate offsets instead of two sample points,thereby ensuring high confidence and significant contribution to the classification.The comparative results on two common benchmark datasets substantiate the superiority of the proposed method compared to existing data generation approaches for few-shot learning,and the effectiveness of the algorithm has been proved through ablation experiments.

A Novel Deep Model with Meta-Learning for Rolling Bearing Few-Shot Fault Diagnosis

Machine learning,especially deep learning,has been highly successful in data-intensive applications;however,the performance of these models will drop significantly when the amount of the training data amount does not meet the requirement.This leads to the so-called few-shot learning(FSL)problem,which requires the model rapidly generalize to new tasks that containing only a few labeled samples.In this paper,we proposed a new deep model,called deep convolutional meta-learning networks,to address the low performance of generalization under limited data for bearing fault diagnosis.The essential of our approach is to learn a base model from the multiple learning tasks using a support dataset and finetune the learnt parameters using few-shot tasks before it can adapt to the new learning task based on limited training data.The proposed method was compared to several FSL methods,including methods with and without pre-training the embedding mapping,and methods with finetuning the classifier or the whole model by utilizing the few-shot data from the target domain.The comparisons are carried out on 1-shot and 10-shot tasks using the Case Western Reserve University bearing dataset and a cylindrical roller bearing dataset.The experimental result illustrates that our method has good performance on the bearing fault diagnosis across various few-shot conditions.In addition,we found that the pretraining process does not always improve the prediction accuracy.

Application of meta-learning in cyberspace security:a survey

In recent years,machine learning has made great progress in intrusion detection,network protection,anomaly detection,and other issues in cyberspace.However,these traditional machine learning algorithms usually require a lot of data to learn and have a low recognition rate for unknown attacks.Among them,“one-shot learning”,“few-shot learning”,and“zero-shot learning”are challenges that cannot be ignored for traditional machine learning.The more intractable problem in cyberspace security is the changeable attack mode.When a new attack mode appears,there are few or even zero samples that can be learned.Meta-learning comes from imitating human problem-solving methods as humans can quickly learn unknown things based on their existing knowledge when learning.Its purpose is to quickly obtain a model with high accuracy and strong generalization through less data training.This article first divides the meta-learning model into five research directions based on different principles of use.They are model-based,metric-based,optimization-based,online-learning-based,or stacked ensemble-based.Then,the current problems in the field of cyberspace security are categorized into three branches:cyber security,information security,and artificial intelligence security according to different perspectives.Then,the application research results of various meta-learning models on these three branches are reviewed.At the same time,based on the characteristics of strong generalization,evolution,and scalability of meta-learning,we contrast and summarize its advantages in solving problems.Finally,the prospect of future deep application of meta-learning in the field of cyberspace security is summarized.

Few-Shot Learning for Discovering Anomalous Behaviors in Edge Networks

Intrusion Detection Systems(IDSs)have a great interest these days to discover complex attack events and protect the critical infrastructures of the Internet of Things(IoT)networks.Existing IDSs based on shallow and deep network architectures demand high computational resources and high volumes of data to establish an adaptive detection engine that discovers new families of attacks from the edge of IoT networks.However,attackers exploit network gateways at the edge using new attacking scenarios(i.e.,zero-day attacks),such as ransomware and Distributed Denial of Service(DDoS)attacks.This paper proposes new IDS based on Few-Shot Deep Learning,named CNN-IDS,which can automatically identify zero-day attacks from the edge of a network and protect its IoT systems.The proposed system comprises two-methodological stages:1)a filtered Information Gain method is to select the most useful features from network data,and 2)one-dimensional Convolutional Neural Network(CNN)algorithm is to recognize new attack types from a network’s edge.The proposed model is trained and validated using two datasets of the UNSW-NB15 and Bot-IoT.The experimental results showed that it enhances about a 3%detection rate and around a 3%–4%falsepositive rate with the UNSW-NB15 dataset and about an 8%detection rate using the BoT-IoT dataset.

Better use of experience from other reservoirs for accurate production forecasting by learn-to-learn method

To assess whether a development strategy will be profitable enough,production forecasting is a crucial and difficult step in the process.The development history of other reservoirs in the same class tends to be studied to make predictions accurate.However,the permeability field,well patterns,and development regime must all be similar for two reservoirs to be considered in the same class.This results in very few available experiences from other reservoirs even though there is a lot of historical information on numerous reservoirs because it is difficult to find such similar reservoirs.This paper proposes a learn-to-learn method,which can better utilize a vast amount of historical data from various reservoirs.Intuitively,the proposed method first learns how to learn samples before directly learning rules in samples.Technically,by utilizing gradients from networks with independent parameters and copied structure in each class of reservoirs,the proposed network obtains the optimal shared initial parameters which are regarded as transferable information across different classes.Based on that,the network is able to predict future production indices for the target reservoir by only training with very limited samples collected from reservoirs in the same class.Two cases further demonstrate its superiority in accuracy to other widely-used network methods.

A memory-friendly class-incremental learning method for hand gesture recognition using HD-sEMG

Hand gesture recognition(HGR)plays a vital role in human-computer interaction.The integration of high-density surface electromyography(HD-sEMG)and deep neural networks(DNNs)has significantly improved the robustness and accuracy of HGR systems.These methods are typically effective for a fixed set of trained gestures.However,the need for new gesture classes over time poses a challenge.Introducing new classes to DNNs can lead to a substantial decrease in accuracy for previously learned tasks,a phenomenon known as“catastrophic forgetting,”especially when the training data for earlier tasks is not retained and retrained.This issue is exacerbated in embedded devices with limited storage,which struggle to store the large-scale data of HD-sEMG.Classincremental learning(CIL)is an effective method to reduce catastrophic forgetting.However,existing CIL methods for HGR rarely focus on reducing memory load.To address this,we propose a memory-friendly CIL method for HGR using HD-sEMG.Our approach includes a lightweight convolutional neural network,named SeparaNet,for feature representation learning,coupled with a nearest-mean-of-exemplars classifier for classifi-cation.We introduce a priority exemplar selection algorithm inspired by the herding effect to maintain a manageable set of exemplars during training.Furthermore,a task-equal-weight exemplar sampling strategy is proposed to effectively reduce memory load while preserving high recognition performance.Experimental results on two datasets demonstrate that our method significantly reduces the number of retained exemplars to only a quarter of that required by other CIL methods,accounting for less than 5%of the total samples,while still achieving comparable average accuracy.

Task-adaptation graph network for few-shot learning

Numerous meta-learning methods focus on the few-shot learning issue,yet most of them assume that various tasks have a shared embedding space,so the generalization ability of the trained model is limited.In order to solve the aforementioned problem,a task-adaptive meta-learning method based on graph neural network(TAGN) is proposed in this paper,where the characterization ability of the original feature extraction network is ameliorated and the classification accuracy is remarkably improved.Firstly,a task-adaptation module based on the self-attention mechanism is employed,where the generalization ability of the model is enhanced on the new task.Secondly,images are classified in non-Euclidean domain,where the disadvantages of poor adaptability of the traditional distance function are overcome.A large number of experiments are conducted and the results show that the proposed methodology has a better performance than traditional task-independent classification methods on two real-word datasets.

Menu Text Recognition of Few-shot Learning

Recent advances in OCR show that end-to-end(E2E)training pipelines including detection and identification can achieve the best results.However,many existing methods usually focus on case insensitive English characters.In this paper,we apply an E2E approach,the multiplex multilingual mask TextSpotter,which performs script recognition at the word level and uses different recognition headers to process different scripts while maintaining uniform loss,thus optimizing script recognition and multiple recognition headers simultaneously.Experiments show that this method is superior to the single-head model with similar number of parameters in endto-end identification tasks.

Few-Shot Graph Classification with Structural-Enhanced Contrastive Learning for Graph Data Copyright Protection

Open-source licenses can promote the development of machine learning by allowing others to access,modify,and redistribute the training dataset.However,not all open-source licenses may be appropriate for data sharing,as some may not provide adequate protections for sensitive or personal information such as social network data.Additionally,some data may be subject to legal or regulatory restrictions that limit its sharing,regardless of the licensing model used.Hence,obtaining large amounts of labeled data can be difficult,time-consuming,or expensive in many real-world scenarios.Few-shot graph classification,as one application of meta-learning in supervised graph learning,aims to classify unseen graph types by only using a small amount of labeled data.However,the current graph neural network methods lack full usage of graph structures on molecular graphs and social network datasets.Since structural features are known to correlate with molecular properties in chemistry,structure information tends to be ignored with sufficient property information provided.Nevertheless,the common binary classification task of chemical compounds is unsuitable in the few-shot setting requiring novel labels.Hence,this paper focuses on the graph classification tasks of a social network,whose complex topology has an uncertain relationship with its nodes'attributes.With two multi-class graph datasets with large node-attribute dimensions constructed to facilitate the research,we propose a novel learning framework that integrates both meta-learning and contrastive learning to enhance the utilization of graph topological information.Extensive experiments demonstrate the competitive performance of our framework respective to other state-of-the-art methods.

Few-shot object detection based on positive-sample improvement

Traditional object detectors based on deep learning rely on plenty of labeled samples,which are expensive to obtain.Few-shot object detection(FSOD)attempts to solve this problem,learning detection objects from a few labeled samples,but the performance is often unsatisfactory due to the scarcity of samples.We believe that the main reasons that restrict the performance of few-shot detectors are:(1)the positive samples is scarce,and(2)the quality of positive samples is low.Therefore,we put forward a novel few-shot object detector based on YOLOv4,starting from both improving the quantity and quality of positive samples.First,we design a hybrid multivariate positive sample augmentation(HMPSA)module to amplify the quantity of positive samples and increase positive sample diversity while suppressing negative samples.Then,we design a selective non-local fusion attention(SNFA)module to help the detector better learn the target features and improve the feature quality of positive samples.Finally,we optimize the loss function to make it more suitable for the task of FSOD.Experimental results on PASCAL VOC and MS COCO demonstrate that our designed few-shot object detector has competitive performance with other state-of-the-art detectors.

Decoupled Two-Phase Framework for Class-Incremental Few-Shot Named Entity Recognition

Class-Incremental Few-Shot Named Entity Recognition(CIFNER)aims to identify entity categories that have appeared with only a few newly added(novel)class examples.However,existing class-incremental methods typically introduce new parameters to adapt to new classes and treat all information equally,resulting in poor generalization.Meanwhile,few-shot methods necessitate samples for all observed classes,making them difficult to transfer into a class-incremental setting.Thus,a decoupled two-phase framework method for the CIFNER task is proposed to address the above issues.The whole task is converted to two separate tasks named Entity Span Detection(ESD)and Entity Class Discrimination(ECD)that leverage parameter-cloning and label-fusion to learn different levels of knowledge separately,such as class-generic knowledge and class-specific knowledge.Moreover,different variants,such as the Conditional Random Field-based(CRF-based),word-pair-based methods in ESD module,and add-based,Natural Language Inference-based(NLI-based)and prompt-based methods in ECD module,are investigated to demonstrate the generalizability of the decoupled framework.Extensive experiments on the three Named Entity Recognition(NER)datasets reveal that our method achieves the state-of-the-art performance in the CIFNER setting.

Recent advances of few-shot learning methods and applications

The rapid development of deep learning provides great convenience for production and life.However,the massive labels required for training models limits further development.Few-shot learning which can obtain a high-performance model by learning few samples in new tasks,providing a solution for many scenarios that lack samples.This paper summarizes few-shot learning algorithms in recent years and proposes a taxonomy.Firstly,we introduce the few-shot learning task and its significance.Secondly,according to different implementation strategies,few-shot learning methods in recent years are divided into five categories,including data augmentation-based methods,metric learning-based methods,parameter optimization-based methods,external memory-based methods,and other approaches.Next,We investigate the application of few-shot learning methods and summarize them from three directions,including computer vision,human-machine language interaction,and robot actions.Finally,we analyze the existing few-shot learning methods by comparing evaluation results on mini Image Net,and summarize the whole paper.

Few-shot node classification via local adaptive discriminant structure learning

Node classification has a wide range of application scenarios such as citation analysis and social network analysis.In many real-world attributed networks,a large portion of classes only contain limited labeled nodes.Most of the existing node classification methods cannot be used for few-shot node classification.To train the model effectively and improve the robustness and reliability of the model with scarce labeled samples,in this paper,we propose a local adaptive discriminant structure learning(LADSL)method for few-shot node classification.LADSL aims to properly represent the nodes in the attributed graphs and learn a metric space with a strong discriminating power by reducing the intra-class variations and enlargingginter-classdifferences.Extensiveexperiments conducted on various attributed networks datasets demonstrate that LADSL is superior to the other methods on few-shot node classification task.

Augmentation-based discriminative meta-learning for cross-machine few-shot fault diagnosis

Deep learning methods have demonstrated promising performance in fault diagnosis tasks.Although the scarcity of data in industrial scenarios limits the practical application of such methods,transfer learning effectively tackles this issue through crossmachine knowledge transfer.Nevertheless,the cross-machine few-shot problem,which is a more general industrial scenario,has been rarely investigated.Existing studies have not considered the cross-machine domain shift problem,which results in poor testing performance.This paper proposes an augmentation-based discriminative meta-learning method to address this issue.In the meta-training process,signal transformation is proposed to increase the meta-task diversity for more robust feature learning,and multi-scale learning is combined for more adaptive feature embedding.In the meta-testing process,limited labeled fault information is used to promote model generalization in the target domain through quasi-meta-training based on data augmentation.Furthermore,a novel hyperbolic prototypical loss is proposed for more discriminative feature representation and separable category prototypes by designing a hyperbolic decision boundary.Cross-machine few-shot diagnosis experiments were conducted using three datasets from different machines,namely,the bearing,motor,and gear datasets.The effectiveness of the proposed method was verified through ablation and comparison studies.

Teachers cooperation:team-knowledge distillation for multiple cross-domain few-shot learning

Although few-shot learning(FSL)has achieved great progress,it is still an enormous challenge especially when the source and target set are from different domains,which is also known as cross-domain few-shot learning(CD-FSL).Utilizing more source domain data is an effective way to improve the performance of CD-FSL.However,knowledge from different source domains may entangle and confuse with each other,which hurts the performance on the target domain.Therefore,we propose team-knowledge distllation networks(TKD-Net)to tackle this problem,which explores a strategy to help the cooperation of multiple teachers.Specifically,we distill knowledge from the cooperation of teacher networks to a single student network in a meta-learning framework.It incorporates task-oriented knowledge distillation and multiple cooperation among teachers to train an efficient student with better generalization ability on unseen tasks.Moreover,our TKD-Net employs both response-based knowledge and relation-based knowledge to transfer more comprehensive and effective knowledge.Extensive experimental results on four fine-grained datasets have demonstrated the effectiveness and superiority of our proposed TKD-Net approach.