Mixup is an effective data augmentation method that generates new augmented samples by aggregating linear combinations of different original samples. However, if there are noises or aberrant features in the original samples, mixup may propagate them to the augmented samples, leading to over-sensitivity of the model to these outliers. To solve this problem, this paper proposes a new mixup method called AMPLIFY. This method uses the attention mechanism of Transformer itself to reduce the influence of noises and aberrant values in the original samples on the prediction results, without increasing additional trainable parameters, and the computational cost is very low, thereby avoiding the problem of high resource consumption in common mixup methods such as Sentence Mixup. The experimental results show that, under a smaller computational resource cost, AMPLIFY outperforms other mixup methods in text classification tasks on seven benchmark datasets, providing new ideas and new ways to further improve the performance of pre-trained models based on the attention mechanism, such as BERT, ALBERT, RoBERTa, and GPT. Our code can be obtained at https://github.com/kiwi-lilo/AMPLIFY.

Chest X-ray (CXR) is an extensively utilized radiological modality for supporting the diagnosis of chest diseases. However, existing research approaches suffer from limitations in effectively integrating multi-scale CXR image features and are also hindered by imbalanced datasets. Therefore, there is a pressing need for further advancement in computer-aided diagnosis (CAD) of thoracic diseases. To tackle these challenges, we propose a multi-branch residual attention network (MBRANet) for thoracic disease diagnosis. MBRANet comprises three components. Firstly, to address the issue of inadequate extraction of spatial and positional information by the convolutional layer, a novel residual structure incorporating a coordinate attention (CA) module is proposed to extract features at multiple scales. Next, based on the concept of a Feature Pyramid Network (FPN), we perform multi-scale feature fusion in the following manner. Thirdly, we propose a novel Multi-Branch Feature Classifier (MFC) approach, which leverages the class-specific residual attention (CSRA) module for classification instead of relying solely on the fully connected layer. In addition, the designed BCEWithLabelSmoothing loss function improves the generalization ability and mitigates the problem of class imbalance by introducing a smoothing factor. We evaluated MBRANet on the ChestX-Ray14, CheXpert, MIMIC-CXR, and IU X-Ray datasets and achieved average AUCs of 0.841, 0.895, 0.805, and 0.745, respectively. Our method outperformed state-of-the-art baselines on these benchmark datasets.

Training with soft labels instead of hard labels can effectively improve the robustness and generalization of deep learning models. Label smoothing often provides uniformly distributed soft labels during the training process, whereas it does not take the semantic difference of labels into account. This article introduces discrimination-aware label smoothing, an adaptive label smoothing approach that learns appropriate distributions of labels for iterative optimization objectives. In this approach, positive and negative samples are employed to provide experience from both sides, and the performances of regularization and model calibration are improved through an iterative learning method. Experiments on five text classification datasets demonstrate the effectiveness of the proposed method.

Deep transfer learning has been widely used to improve the versatility of models. In the problem of cross-domain fault diagnosis in rolling bearings, most models require that the given data have a similar distribution, which limits the diagnostic effect and generalization of the model. This paper proposes a deep reconstruction transfer convolutional neural network (DRTCNN), which satisfies the domain adaptability of the model under cross-domain conditions. Firstly, the model uses a deep reconstruction convolutional automatic encoder for feature extraction and data reconstruction. Through sharing parameters and unsupervised training, the structural information of target domain samples is effectively used to extract domain-invariant features. Secondly, a new subdomain alignment loss function is introduced to align the subdomain distribution of the source domain and the target domain, which can improve the classification accuracy by reducing the intra-class distance and increasing the inter-class distance. In addition, a label smoothing algorithm considering the credibility of the sample is introduced to train the model classifier to avoid the impact of wrong labels on the training process. Three datasets are used to verify the versatility of the model, and the results show that the model has a high accuracy and stability.

Machine learning-based algorithms demonstrate impressive performance across numerous fields; however, they continue to suffer from certain limitations. Even sophisticated and precise algorithms often make erroneous predictions when implemented with datasets having different distributions compared to the training set. Out-of-distribution (OOD) detection, which distinguishes data with different distributions from that of the training set, is a critical research area necessary to overcome these limitations and create more reliable algorithms. The OOD issue, particularly concerning image data, has been extensively studied. However, recently developed OOD methods do not fulfill the expectation that OOD performance will increase as the accuracy of in-distribution classification improves. Our research presents a comprehensive study on OOD detection performance across multiple models and training methodologies to verify this phenomenon. Specifically, we explore various pre-trained models popular in the computer vision field with both old and new OOD detection methods. The experimental results highlight the performance disparity in existing OOD methods. Based on these observations, we introduce Trimmed Rank with Inverse softMax probability (TRIM), a remarkably simple yet effective method for model weights with newly developed training methods. The proposed method could serve as a potential tool for enhancing OOD detection performance owing to its promising results. The OOD performance of TRIM is highly compatible with the in-distribution accuracy model and may bridge the efforts on improving in-distribution accuracy to the ability to distinguish OOD data.

Multi-view deep neural networks have shown excellent performance on 3D shape classification tasks. However, global features aggregated from multiple views data often lack content information and spatial relationship, which leads to difficult identification the small variance among subcategories in the same category. To solve this problem, in this paper, a novel multiscale dilated convolution neural network termed as MSDCNN is proposed for multi-view fine-grained 3D shape classification. Firstly, a sequence of views are rendered from 12-viewpoints around the input 3D shape by the sequential view capturing module. Then, the first 22 convolution layers of ResNeXt50 is employed to extract the semantic features of each view, and a global mixed feature map is obtained through the element-wise maximum operation of the 12 output feature maps. Furthermore, attention dilated module (ADM), which combines four concatenated attention dilated block (ADB), is designed to extract larger receptive field features from global mixed feature map to enhance context information among the views. Specifically, each ADB is consisted by an attention mechanism module and a dilated convolution with different dilation rates. In addition, prediction module with label smoothing is proposed to classify features, which contains 3 × 3 convolution and adaptive average pooling. The performance of our method is validated experimentally on the ModelNet10, ModelNet40 and FG3D datasets. Experimental results demonstrate the effectiveness and superiority of the proposed MSDCNN framework for 3D shape fine-grained classification.

Recent research on semi-supervised learning (SSL) is mainly based on the method of consistency regularization, which relies on domain-specific data augmentation. Pseudo-labeling is a more general method that has no such restrictions but performs limited by noisy training. We combine both approaches and focus on generating pseudo-labels using domain-independent weak augmentation. In this article, we propose ReFixMatch-LS and apply it to the classification of medical images. First, we reduce the impact of noisy artificial labels by label smoothing and consistent regularization. Then, by recording high-confidence pseudo-labels generated from each epoch during training, we reuse the generated pseudo-labels to train the model in the subsequent epochs. ReFixMatch-LS effectively increases the number of pseudo-labels and improves the model performance. We validate the effectiveness of ReFixMatch-LS on skin lesion diagnosis in the ISIC 2018 and ISIC 2019 challenge datasets, obtaining AUCs of 91.54%, 93.68%, 94.55%, and 95.47% on the four proportions of labeled data from ISIC 2018.

Head pose estimation is one of the essential tasks in computer vision, which predicts the Euler angles of the head in an image. In recent years, CNN-based methods for head pose estimation have achieved excellent performance. Their training relies on RGB images providing facial landmarks or depth images from RGBD cameras. However, labeling facial landmarks is complex for large angular head poses in RGB images, and RGBD cameras are unsuitable for outdoor scenes. We propose a simple and effective annotation method for the head pose in RGB images. The novelty method uses a 3D virtual human head to simulate the head pose in the RGB image. The Euler angle can be calculated from the change in coordinates of the 3D virtual head. We then create a dataset using our annotation method: 2DHeadPose dataset, which contains a rich set of attributes, dimensions, and angles. Finally, we propose Gaussian label smoothing to suppress annotation noises and reflect inter-class relationships. A baseline approach is established using Gaussian label smoothing. Experiments demonstrate that our annotation method, datasets, and Gaussian label smoothing are very effective. Our baseline approach surpasses most current state-of-the-art methods. The annotation tool, dataset, and source code are publicly available at https://github.com/youngnuaa/2DHeadPose.

In recent times, many studies concerning surgical video analysis are being conducted due to its growing importance in many medical applications. In particular, it is very important to be able to recognize the current surgical phase because the phase information can be utilized in various ways both during and after surgery. This paper proposes an efficient phase recognition network, called MomentNet, for cholecystectomy endoscopic videos. Unlike LSTM-based network, MomentNet is based on a multi-stage temporal convolutional network. Besides, to improve the phase prediction accuracy, the proposed method adopts a new loss function to supplement the general cross entropy loss function. The new loss function significantly improves the performance of the phase recognition network by constraining un-desirable phase transition and preventing over-segmentation. In addition, MomnetNet effectively applies positional encoding techniques, which are commonly applied in transformer architectures, to the multi-stage temporal convolution network. By using the positional encoding techniques, MomentNet can provide important temporal context, resulting in higher phase prediction accuracy. Furthermore, the MomentNet applies label smoothing technique to suppress overfitting and replaces the backbone network for feature extraction to further improve the network performance. As a result, the MomentNet achieves 92.31% accuracy in the phase recognition task with the Cholec80 dataset, which is 4.55% higher than that of the baseline architecture.

White blood cells (WBCs) in the human immune system defend against infection and protect the body from external hazardous objects. They are comprised of neutrophils, eosinophils, basophils, monocytes, and lymphocytes, whereby each accounts for a distinct percentage and performs specific functions. Traditionally, the clinical laboratory procedure for quantifying the specific types of white blood cells is an integral part of a complete blood count (CBC) test, which aids in monitoring the health of people. With the advancements in deep learning, blood film images can be classified in less time and with high accuracy using various algorithms. This paper exploits a number of state-of-the-art deep learning models and their variations based on CNN architecture. A comparative study on model performance based on accuracy, F1-score, recall, precision, number of parameters, and time was conducted, and DenseNet161 was found to demonstrate a superior performance among its counterparts. In addition, advanced optimization techniques such as normalization, mixed-up augmentation, and label smoothing were also employed on DenseNet to further refine its performance.