Neurocomputing最新文献

Few-shot learning (FSL) confronts notable challenges due to the disparity between training and testing categories, leading to channel bias in neural networks and hindering accurate feature discernment. To address this, we introduce Biased-Reduction Attentive Network (BRAVE), an innovative model that incorporates a refined Vector Quantized Variational Autoencoder (VQ-VAE) backbone, enhanced with our Diverse Quantization (DQ) Module, for unbiased, fine-grained feature creation. Alongside, our Sample Attention (SA) Module is utilized for extracting discriminative features from these unbiased, fine-grained features. The DQ Module in BRAVE strategically integrates prior distribution regularization and stochastic masking with Gumbel sampling for balanced and diverse codebook engagement, while the SA Module leverages inter-sample dynamics for identifying critical features. This synergy effectively counters channel bias and boosts classification accuracy in FSL setups, surpassing current leading methods. Our approach represents a practical balance between preserving detailed features through the decoder and ensuring classification effectiveness, marking a significant advance in FSL. BRAVE’s implementation is accessible for community use and further exploration. Code and models available at https://github.com/ApocalypsezZ/BRAVE.

Graph computation via Graph Neural Networks (GNNs) is emerging as a pivotal approach for addressing the challenges in image classification tasks. This paper introduces a novel strategy for image classification using minimal labeled data from the mini-ImageNet database. The primary contributions include the development of an innovative Fine-Grained Feature Descriptor (FGFD) module. Following this, the GNN is employed at a more granular level to enhance image classification efficiency. Additionally, ablation studies were conducted in conjunction with existing state-of-the-art systems for few-shot image classification. Comparative analyses were performed, and the simulation results demonstrate that the proposed method significantly improves classification accuracy over traditional few-shot image classification methods.

Transformer-based language models have achieved significant success in various domains. However, the data-intensive nature of the transformer architecture requires much labeled data, which is challenging in low-resource scenarios (i.e., few-shot learning (FSL)). The main challenge of FSL is the difficulty of training robust models on small amounts of samples, which frequently leads to overfitting. Here we present Mask-BERT, a simple and modular framework to help BERT-based architectures tackle FSL. The proposed approach fundamentally differs from existing FSL strategies such as prompt tuning and meta-learning. The core idea is to selectively apply masks on text inputs and filter out irrelevant information, which guides the model to focus on discriminative tokens that influence prediction results. In addition, to make the text representations from different categories more separable and the text representations from the same category more compact, we introduce a contrastive learning loss function. Experimental results on open-domain and medical-domain datasets demonstrate the effectiveness of Mask-BERT. Code and data are available at: github.com/WenxiongLiao/mask-bert

As a primary optimization method for neural networks, gradient descent algorithm has received significant attention in the recent development of deep neural networks. However, current gradient descent algorithms still suffer from drawbacks such as an excess of hyperparameters, getting stuck in local optima, and poor generalization. This paper introduces a novel Caputo fractional-order gradient descent (MFFGD) algorithm to address these limitations. It provides fractional-order gradient derivation and error analysis for different activation functions and loss functions within the network, simplifying the computation of traditional fractional order gradients. Additionally, by introducing a memory factor to record past gradient variations, MFFGD achieves adaptive adjustment capabilities. Comparative experiments were conducted on multiple sets of datasets with different modalities, and the results, along with theoretical analysis, demonstrate the superiority of MFFGD over other optimizers.

Organs-at-risk (OARs) segmentation in computed tomography (CT) is a fundamental step in the radiotherapy workflow, which has been prone as a time-consuming and labor-intensive task. Deep neural networks (DNNs) have gained significant popularity in the field of OAR segmentation tasks, achieving remarkable progress in clinical practice. Typically, OARs are distributed throughout different areas of the body and require varying thicknesses of CT scans for better diagnosis and segmentation in clinical. Most DNN-based segmentation focuses on single-thickness CT scans, limiting their applicability to varying thicknesses due to a lack of diverse thickness-related feature learning. While pre-training with the denoising diffusion probabilistic model (DDPM) offers an effective solution for dense feature learning, current works are constrained in addressing feature diversity, as exemplified by scenarios such as multi-thickness CT. To address the above challenges, this paper introduces a novel pre-training approach called DiffMT. This approach leverages the DDPM to extract valuable features from multi-thickness CT images. By transferring the pre-trained DDPM to the downstream segmentation for fine-tuning, the model gains proficiency in learning diverse multi-thickness CT features, leading to precise segmentation across varied thicknesses. We explore DiffMT’s feature learning capacity through experiments involving pre-trained models of varying sizes and different denoising thicknesses. Subsequently, thorough experiments comparing DDPM-based segmentation with other state-of-the-art (SOTA) CT segmentation methods, along with assessments on diverse OARs and modalities, empirically demonstrate that the proposed DiffMT method outperforms the control methods. The codes are available at https://github.com/ychengrong/DiffMT.

Haze is one of the atmospheric image degradations that causes severe distortions to outdoor images such as low contrast, color shift, and structure damage. Due to the unique physical characteristics of haze, the quality of hazy images is not accurately assessed using general-purpose image quality assessment (IQA) approaches. Therefore, several haze-aware IQA approaches have been proposed to provide more efficient dehazing quality evaluation. These approaches extract several haze-aware features to be either combined to form a single IQA metric or fed to a regression model that predicts the dehazing quality. However, these haze-relevant features are extracted using pixel intensity, in which luminance and structure information are inseparable, leading to less correlation between such features and the type of degradation they are supposed to represent. To address this issue, we propose a singular value decomposition (SVD) based IQA metric that can effectively separate the luminance component of an image from structure. This separation offers the ability to accurately evaluate the degradation at two different levels i.e. luminance and structure. The experimental results show that our proposed SVD-based dehazing quality evaluator (SDQE) outperforms the existing state-of-the-art non-reference IQA metrics in terms of accuracy and processing time.

Image dehazing has become a necessary area of research with the increasing popularity and demand of computer vision systems. Image dehazing is a method to remove haze from an image to improve its visual quality. Dehazing techniques are widely employed in a variety of computer vision applications to enhance their overall performance. Many techniques have been proposed by researchers in recent years to eliminate the haze from an image. However, there is a lack of available literature that provides a summary of deep learning-based state-of-the-art image dehazing methods. In this study, we provide a detailed review of recently proposed image dehazing techniques based on deep neural networks such as CNN, GAN, RNN, RCNN, and Transformer. A concise review of significant applications of image dehazing, benchmark datasets, and various performance metrics are also presented. We compare the state-of-the-art methods quantitatively using performance evaluation metrics such as SSIM and PSNR. Finally, this study discusses the fundamental difficulties associated with image dehazing approaches that need to be further explored.

For deep learning-based industrial anomaly detection, it is still challenging to get adequate images for model training and achieve cold start for cross-product migration, restricting their practical application in real industrial production. Herein, an innovative few-shot anomaly detection network DMMGNet based on discrimination mapping and memory bank mean guidance strategies are demonstrated, which is trained by a new two-branch data augmentation technique. By separating the features stored in memory bank from the features used for training, the two-branch data augmentation method can significantly improve the robustness of few-shot model training and reduce the redundance of memory bank. In the elaborately designed discrimination mapping module, new negative samples are generated by adding dynamic Gaussian noise to normal samples along the channel dimension in feature space to solve the problem of sample imbalance. Meanwhile, the discrimination mapping module also helps to map the feature distribution of positive samples to the target domain more efficiently and reduce the deviation of feature domain, conducive to a more precise separation of positive and negative samples. In addition, a novel mean guidance approach with an optimized loss function is developed to guide the positive sample feature mapping by specifying the local feature space center to form a clear feature domain contour and enhance the detection accuracy. The multiple experimental results validate that our DMMGNet outperforms the most advanced anomaly detection counterparts on image-level AUROC, showing an increase by 0.3–3 % on both MVTec AD and MPDD benchmarks under several few-shot scenarios.

Accurately segmenting and staging tumor lesions in cancer patients presents a significant challenge for radiologists, but it is essential for devising effective treatment plans including radiation therapy, personalized medicine, and surgical options. The integration of artificial intelligence (AI), particularly deep learning (DL), has become a useful tool for radiologists, enhancing their ability to understand tumor biology and deliver personalized care to patients with H&N tumors. Segmenting H&N tumor lesions using Positron Emission Tomography/Computed Tomography (PET/CT) images has gained significant attention. However, the diverse shapes and sizes of tumors, along with indistinct boundaries between malignant and normal tissues, present significant challenges in effectively fusing PET and CT images. To overcome these challenges, various DL-based models have been developed for segmenting tumor lesions in PET/CT images. This article reviews multimodality (PET/CT) based H&N tumor lesions segmentation methods. We firstly discuss the strengths and limitations of PET/CT imaging and the importance of DL-based models in H&N tumor lesion segmentation. Second, we examine the current state-of-the-art DL models for H&N tumor segmentation, categorizing them into UNet, VNet, Vision Transformer, and miscellaneous models based on their architectures. Third, we explore the annotation and evaluation processes, addressing challenges in segmentation annotation and discussing the metrics used to assess model performance. Finally, we discuss several open challenges and provide some avenues for future research in H&N tumor lesion segmentation.

Spiking neural network (SNN), as a brain-inspired model, possesses outstanding low power consumption and the ability to mimic biological neuron mechanisms. Embodied vision is a promising field, which requires the low-power advantage of SNNs. However, SNN faces difficulties in achieving real-time, high generalization ability, and robustness in the implementation of embodied vision due to the limitations of existing training methods. In this paper, to prevent model overfitting to noise and unknown environment in embodied neuromorphic visual intelligence, we present a new and efficient learning strategy designed to enhance the training performance of deep SNNs, called Spiking Maximum Entropy Intrinsic Learning (SMEIL). The learning algorithm essentially promotes the perturbation of the underlying source distribution, which in turn enlarges the predictive uncertainty of the current model. This approach enhances the model's robustness and improves its ability to generalize during the training process. Superior performance across a variety of data sets is achieved, and different types of noise are added to SMEIL algorithm for testing its robustness. Experiments show that SMEIL can consistently improve the learning robustness over each noise disturbance, and can cut down the power consumption during training significantly. Hence, it is a powerful method for advancing direct training of deep SNNs, and opens a novel point of view for developing novel spike-based learning algorithm towards embodied neuromorphic intelligence.