Feature-wise scaling and shifting: Improving the generalization capability of neural networks through capturing independent information of features.

From the perspective of input features, information can be divided into independent information and correlation information. Current neural networks mainly concentrate on the capturing of correlation information through connection weight parameters supplemented by bias parameters. This paper introduces feature-wise scaling and shifting (FwSS) into neural networks for capturing independent information of features, and proposes a new neural network FwSSNet. In the network, a pair of scale and shift parameters is added before each input of each network layer, and bias is removed. The parameters are initialized as 1 and 0, respectively, and trained at separate learning rates, to guarantee the fully capturing of independence and correlation information. The learning rates of FwSS parameters depend on input data and the training speed ratios of adjacent FwSS and connection sublayers, meanwhile those of weight parameters remain unchanged as plain networks. Further, FwSS unifies the scaling and shifting operations in batch normalization (BN), and FwSSNet with BN is established through introducing a preprocessing layer. FwSS parameters except those in the last layer of the network can be simply trained at the same learning rate as weight parameters. Experiments show that FwSS is generally helpful in improving the generalization capability of both fully connected neural networks and deep convolutional neural networks, and FWSSNets achieve higher accuracies on UCI repository and CIFAR-10.

Complementary label learning based on knowledge distillation.

Complementary label learning (CLL) is a type of weakly supervised learning method that utilizes the category of samples that do not belong to a certain class to learn their true category. However, current CLL methods mainly rely on rewriting classification losses without fully leveraging the supervisory information in complementary labels. Therefore, enhancing the supervised information in complementary labels is a promising approach to improve the performance of CLL. In this paper, we propose a novel framework called Complementary Label Enhancement based on Knowledge Distillation (KDCL) to address the lack of attention given to complementary labels. KDCL consists of two deep neural networks: a teacher model and a student model. The teacher model focuses on softening complementary labels to enrich the supervision information in them, while the student model learns from the complementary labels that have been softened by the teacher model. Both the teacher and student models are trained on the dataset that contains only complementary labels. To evaluate the effectiveness of KDCL, we conducted experiments on four datasets, namely MNIST, F-MNIST, K-MNIST and CIFAR-10, using two sets of teacher-student models (Lenet-5+MLP and DenseNet-121+ResNet-18) and three CLL algorithms (PC, FWD and SCL-NL). Our experimental results demonstrate that models optimized by KDCL outperform those trained only with complementary labels in terms of accuracy.

Class-imbalanced complementary-label learning via weighted loss.

Complementary-label learning (CLL) is widely used in weakly supervised classification, but it faces a significant challenge in real-world datasets when confronted with class-imbalanced training samples. In such scenarios, the number of samples in one class is considerably lower than in other classes, which consequently leads to a decline in the accuracy of predictions. Unfortunately, existing CLL approaches have not investigate this problem. To alleviate this challenge, we propose a novel problem setting that enables learning from class-imbalanced complementary labels for multi-class classification. To tackle this problem, we propose a novel CLL approach called Weighted Complementary-Label Learning (WCLL). The proposed method models a weighted empirical risk minimization loss by utilizing the class-imbalanced complementary labels, which is also applicable to multi-class imbalanced training samples. Furthermore, we derive an estimation error bound to provide theoretical assurance. To evaluate our approach, we conduct extensive experiments on several widely-used benchmark datasets and a real-world dataset, and compare our method with existing state-of-the-art methods. The proposed approach shows significant improvement in these datasets, even in the case of multiple class-imbalanced scenarios. Notably, the proposed method not only utilizes complementary labels to train a classifier but also solves the problem of class imbalance.

MTRT: Motion Trajectory Reconstruction Transformer for EEG-Based BCI Decoding.

Brain computer interface (BCI) is a system that directly uses brain neural activities to communicate with the outside world. Recently, the decoding of the human upper limb based on electroencephalogram (EEG) signals has become an important research branch of BCI. Even though existing research models are capable of decoding upper limb trajectories, the performance needs to be improved to make them more practical for real-world applications. This study is attempt to reconstruct the continuous and nonlinear multi-directional upper limb trajectory based on Chinese sign language. Here, to reconstruct the upper limb motion trajectory effectively, we propose a novel Motion Trajectory Reconstruction Transformer (MTRT) neural network that utilizes the geometric information of human joint points and EEG neural activity signals to decode the upper limb trajectory. Specifically, we use human upper limb bone geometry properties as reconstruction constraints to obtain more accurate trajectory information of the human upper limbs. Furthermore, we propose a MTRT neural network based on this constraint, which uses the shoulder, elbow, and wrist joint point information and EEG signals of brain neural activity during upper limb movement to train its parameters. To validate the model, we collected the synchronization information of EEG signals and upper limb motion joint points of 20 subjects. The experimental results show that the reconstruction model can accurately reconstruct the motion trajectory of the shoulder, elbow, and wrist of the upper limb, achieving superior performance than the compared methods. This research is very meaningful to decode the limb motion parameters for BCI, and it is inspiring for the motion decoding of other limbs and other joints.

FAPI-Net: A lightweight interpretable network based on feature augmentation and prototype interpretation.

With the increasing application of deep neural networks, their performance requirements in various fields are increasing. Deep neural network models with higher performance generally have a high number of parameters and computation (FLOPs, Floating Point Operations), and have the black-box characteristic. This hinders the deployment of deep neural network models on low-power platforms, as well as sustainable development in high-risk decision-making fields. However, there is little work to ensure the interpretability of the model in the research on the lightweight of the deep neural network model. This paper proposed FAPI-Net (feature augmentation and prototype interpretation), a lightweight interpretable network. It combined feature augmentation convolution blocks and the prototype dictionary interpretability (PDI) module. The feature augmentation convolution block is composed of lightweight feature-map augmentation (FA) modules and a residual connection stack. The FA module could effectively reduce network parameters and computation without losing network accuracy. The PDI module can realize the visualization of model classification reasoning. FAPI-Net is designed regarding MobileNetV3's structure, and our experiments show that the FAPI-Net is more effective than MobileNetV3 and other advanced lightweight CNNs. Params and FLOPs on the ILSVRC2012 dataset are 2 and 20% lower than that on MobileNetV3, respectively, and FAPI-Net with a trainable PDI module has almost no loss of accuracy compared with baseline models. In addition, the ablation experiment on the CIFAR-10 dataset proved the effectiveness of the FA module used in FAPI-Net. The decision reasoning visualization experiments show that FAPI-Net could make the classification decision process of specific test images transparent.

Low-Dose CT Denoising via Sinogram Inner-Structure Transformer.

Low-Dose Computed Tomography (LDCT) technique, which reduces the radiation harm to human bodies, is now attracting increasing interest in the medical imaging field. As the image quality is degraded by low dose radiation, LDCT exams require specialized reconstruction methods or denoising algorithms. However, most of the recent effective methods overlook the inner-structure of the original projection data (sinogram) which limits their denoising ability. The inner-structure of the sinogram represents special characteristics of the data in the sinogram domain. By maintaining this structure while denoising, the noise can be obviously restrained. Therefore, we propose an LDCT denoising network namely Sinogram Inner-Structure Transformer (SIST) to reduce the noise by utilizing the inner-structure in the sinogram domain. Specifically, we study the CT imaging mechanism and statistical characteristics of sinogram to design the sinogram inner-structure loss including the global and local inner-structure for restoring high-quality CT images. Besides, we propose a sinogram transformer module to better extract sinogram features. The transformer architecture using a self-attention mechanism can exploit interrelations between projections of different view angles, which achieves an outstanding performance in sinogram denoising. Furthermore, in order to improve the performance in the image domain, we propose the image reconstruction module to complementarily denoise both in the sinogram and image domain.

Active learning for efficient analysis of high-throughput nanopore data.

MOTIVATION: As the third-generation sequencing technology, nanopore sequencing has been used for high-throughput sequencing of DNA, RNA, and even proteins. Recently, many studies have begun to use machine learning technology to analyze the enormous data generated by nanopores. Unfortunately, the success of this technology is due to the extensive labeled data, which often suffer from enormous labor costs. Therefore, there is an urgent need for a novel technology that can not only rapidly analyze nanopore data with high-throughput, but also significantly reduce the cost of labeling. To achieve the above goals, we introduce active learning to alleviate the enormous labor costs by selecting the samples that need to be labeled. This work applies several advanced active learning technologies to the nanopore data, including the RNA classification dataset (RNA-CD) and the Oxford Nanopore Technologies barcode dataset (ONT-BD). Due to the complexity of the nanopore data (with noise sequence), the bias constraint is introduced to improve the sample selection strategy in active learning. Results: The experimental results show that for the same performance metric, 50% labeling amount can achieve the best baseline performance for ONT-BD, while only 15% labeling amount can achieve the best baseline performance for RNA-CD. Crucially, the experiments show that active learning technology can assist experts in labeling samples, and significantly reduce the labeling cost. Active learning can greatly reduce the dilemma of difficult labeling of high-capacity nanopore data. We hope active learning can be applied to other problems in nanopore sequence analysis. AVAILABILITY AND IMPLEMENTATION: The main program is available at https://github.com/guanxiaoyu11/AL-for-nanopore. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Uncertainty Ordinal Multi-Instance Learning for Breast Cancer Diagnosis.

Ordinal multi-instance learning (OMIL) deals with the weak supervision scenario wherein instances in each training bag are not only multi-class but also have rank order relationships between classes, such as breast cancer, which has become one of the most frequent diseases in women. Most of the existing work has generally been to classify the region of interest (mass or microcalcification) on the mammogram as either benign or malignant, while ignoring the normal mammogram classification. Early screening for breast disease is particularly important for further diagnosis. Since early benign lesion areas on a mammogram are very similar to normal tissue, three classifications of mammograms for the improved screening of early benign lesions are necessary. In OMIL, an expert will only label the set of instances (bag), instead of labeling every instance. When labeling efforts are focused on the class of bags, ordinal classes of the instance inside the bag are not labeled. However, recent work on ordinal multi-instance has used the traditional support vector machine to solve the multi-classification problem without utilizing the ordinal information regarding the instances in the bag. In this paper, we propose a method that explicitly models the ordinal class information for bags and instances in bags. Specifically, we specify a key instance from the bag as a positive instance of bags, and design ordinal minimum uncertainty loss to iteratively optimize the selected key instances from the bags. The extensive experimental results clearly prove the effectiveness of the proposed ordinal instance-learning approach, which achieves 52.021% accuracy, 61.471% sensitivity, 47.206% specificity, 57.895% precision, and an 59.629% F1 score on a DDSM dataset.

Automated Bone Age Assessment: A New Three-Stage Assessment Method from Coarse to Fine.

Bone age assessment (BAA) based on X-ray imaging of the left hand and wrist can accurately reflect the degree of the body's physiological development and physical condition. However, the traditional manual evaluation method relies too much on inefficient specialist labor. In this paper, to propose automatic BAA, we introduce a hierarchical convolutional neural network to detect the regions of interest (ROI) and classify the bone grade. Firstly, we establish a dataset of children's BAA containing 2518 left hand X-rays. Then, we use the fine-grained classification to obtain the grade of the region of interest via object detection. Specifically, fine-grained classifiers are based on context-aware attention pooling (CAP). Finally, we perform the model assessment of bone age using the third version of the Tanner-Whitehouse (TW3) methodology. The end-to-end BAA system provides bone age values, the detection results of 13 ROIs, and the bone maturity of the ROIs, which are convenient for doctors to obtain information for operation. Experimental results on the public dataset and clinical dataset show that the performance of the proposed method is competitive. The accuracy of bone grading is 86.93%, and the mean absolute error (MAE) of bone age is 7.68 months on the clinical dataset. On public dataset, the MAE is 6.53 months. The proposed method achieves good performance in bone age assessment and is superior to existing fine-grained image classification methods.

InfoAT: Improving Adversarial Training Using the Information Bottleneck Principle.

Adversarial training (AT) has shown excellent high performance in defending against adversarial examples. Recent studies demonstrate that examples are not equally important to the final robustness of models during AT, that is, the so-called hard examples that can be attacked easily exhibit more influence than robust examples on the final robustness. Therefore, guaranteeing the robustness of hard examples is crucial for improving the final robustness of the model. However, defining effective heuristics to search for hard examples is still difficult. In this article, inspired by the information bottleneck (IB) principle, we uncover that an example with high mutual information of the input and its associated latent representation is more likely to be attacked. Based on this observation, we propose a novel and effective adversarial training method (InfoAT). InfoAT is encouraged to find examples with high mutual information and exploit them efficiently to improve the final robustness of models. Experimental results show that InfoAT achieves the best robustness among different datasets and models in comparison with several state-of-the-art methods.

S2Snet: deep learning for low molecular weight RNA identification with nanopore.

Ribonucleic acid (RNA) is a pivotal nucleic acid that plays a crucial role in regulating many biological activities. Recently, one study utilized a machine learning algorithm to automatically classify RNA structural events generated by a Mycobacterium smegmatis porin A nanopore trap. Although it can achieve desirable classification results, compared with deep learning (DL) methods, this classic machine learning requires domain knowledge to manually extract features, which is sophisticated, labor-intensive and time-consuming. Meanwhile, the generated original RNA structural events are not strictly equal in length, which is incompatible with the input requirements of DL models. To alleviate this issue, we propose a sequence-to-sequence (S2S) module that transforms the unequal length sequence (UELS) to the equal length sequence. Furthermore, to automatically extract features from the RNA structural events, we propose a sequence-to-sequence neural network based on DL. In addition, we add an attention mechanism to capture vital information for classification, such as dwell time and blockage amplitude. Through quantitative and qualitative analysis, the experimental results have achieved about a 2% performance increase (accuracy) compared to the previous method. The proposed method can also be applied to other nanopore platforms, such as the famous Oxford nanopore. It is worth noting that the proposed method is not only aimed at pursuing state-of-the-art performance but also provides an overall idea to process nanopore data with UELS.

Neural Decoding of Chinese Sign Language With Machine Learning for Brain-Computer Interfaces.

Limb motion decoding is an important part of brain-computer interface (BCI) research. Among the limb motion, sign language not only contains rich semantic information and abundant maneuverable actions but also provides different executable commands. However, many researchers focus on decoding the gross motor skills, such as the decoding of ordinary motor imagery or simple upper limb movements. Here we explored the neural features and decoding of Chinese sign language from electroencephalograph (EEG) signal with motor imagery and motor execution. Sign language not only contains rich semantic information, but also has abundant maneuverable actions, and provides us with more different executable commands. In this paper, twenty subjects were instructed to perform movement execution and movement imagery based on Chinese sign language. Seven classifiers are employed to classify the selected features of sign language EEG. L1 regularization is used to learn and select features that contain more information from the mean, power spectral density, sample entropy, and brain network connectivity. The best average classification accuracy of the classifier is 89.90% (imagery sign language is 83.40%). These results have shown the feasibility of decoding between different sign languages. The source location reveals that the neural circuits involved in sign language are related to the visual contact area and the pre-movement area. Experimental evaluation shows that the proposed decoding strategy based on sign language can obtain outstanding classification results, which provides a certain reference value for the subsequent research of limb decoding based on sign language.

GACDN: generative adversarial feature completion and diagnosis network for COVID-19.

BACKGROUND: The outbreak of coronavirus disease 2019 (COVID-19) causes tens of million infection world-wide. Many machine learning methods have been proposed for the computer-aided diagnosis between COVID-19 and community-acquired pneumonia (CAP) from chest computed tomography (CT) images. Most of these methods utilized the location-specific handcrafted features based on the segmentation results to improve the diagnose performance. However, the prerequisite segmentation step is time-consuming and needs the intervention by lots of expert radiologists, which cannot be achieved in the areas with limited medical resources. METHODS: We propose a generative adversarial feature completion and diagnosis network (GACDN) that simultaneously generates handcrafted features by radiomic counterparts and makes accurate diagnoses based on both original and generated features. Specifically, we first calculate the radiomic features from the CT images. Then, in order to fast obtain the location-specific handcrafted features, we use the proposed GACDN to generate them by its corresponding radiomic features. Finally, we use both radiomic features and location-specific handcrafted features for COVID-19 diagnosis. RESULTS: For the performance of our generated location-specific handcrafted features, the results of four basic classifiers show that it has an average of 3.21% increase in diagnoses accuracy. Besides, the experimental results on COVID-19 dataset show that our proposed method achieved superior performance in COVID-19 vs. community acquired pneumonia (CAP) classification compared with the state-of-the-art methods. CONCLUSIONS: The proposed method significantly improves the diagnoses accuracy of COVID-19 vs. CAP in the condition of incomplete location-specific handcrafted features. Besides, it is also applicable in some regions lacking of expert radiologists and high-performance computing resources.

Towards evaluating the robustness of deep diagnostic models by adversarial attack.

Deep learning models (with neural networks) have been widely used in challenging tasks such as computer-aided disease diagnosis based on medical images. Recent studies have shown deep diagnostic models may not be robust in the inference process and may pose severe security concerns in clinical practice. Among all the factors that make the model not robust, the most serious one is adversarial examples. The so-called "adversarial example" is a well-designed perturbation that is not easily perceived by humans but results in a false output of deep diagnostic models with high confidence. In this paper, we evaluate the robustness of deep diagnostic models by adversarial attack. Specifically, we have performed two types of adversarial attacks to three deep diagnostic models in both single-label and multi-label classification tasks, and found that these models are not reliable when attacked by adversarial example. We have further explored how adversarial examples attack the models, by analyzing their quantitative classification results, intermediate features, discriminability of features and correlation of estimated labels for both original/clean images and those adversarial ones. We have also designed two new defense methods to handle adversarial examples in deep diagnostic models, i.e., Multi-Perturbations Adversarial Training (MPAdvT) and Misclassification-Aware Adversarial Training (MAAdvT). The experimental results have shown that the use of defense methods can significantly improve the robustness of deep diagnostic models against adversarial attacks.