Neural Networks最新文献

Vision Transformer have achieved impressive performance in image super-resolution. However, they suffer from low inference speed mainly because of the quadratic complexity of multi-head self-attention (MHSA), which is the key to learning long-range dependencies. On the contrary, most CNN-based methods neglect the important effect of global contextual information, resulting in inaccurate and blurring details. If one can make the best of both Transformers and CNNs, it will achieve a better trade-off between image quality and inference speed. Based on this observation, firstly assume that the main factor affecting the performance in the Transformer-based SR models is the general architecture design, not the specific MHSA component. To verify this, some ablation studies are made by replacing MHSA with large kernel convolutions, alongside other essential module replacements. Surprisingly, the derived models achieve competitive performance. Therefore, a general architecture design GlobalSR is extracted by not specifying the core modules including blocks and domain embeddings of Transformer-based SR models. It also contains three practical guidelines for designing a lightweight SR network utilizing image-level global contextual information to reconstruct SR images. Following the guidelines, the blocks and domain embeddings of GlobalSR are instantiated via Deformable Convolution Attention Block (DCAB) and Fast Fourier Convolution Domain Embedding (FCDE), respectively. The instantiation of general architecture, termed GlobalSR-DF, proposes a DCA to extract the global contextual feature by utilizing Deformable Convolution and a Hadamard product as the attention map at the block level. Meanwhile, the FCDE utilizes the Fast Fourier to transform the input spatial feature into frequency space and then extract image-level global information from it by convolutions. Extensive experiments demonstrate that GlobalSR is the key part in achieving a superior trade-off between SR quality and efficiency. Specifically, our proposed GlobalSR-DF outperforms state-of-the-art CNN-based and ViT-based SISR models regarding accuracy-speed trade-offs with sharp and natural details.

In unsupervised domain adaptive object detection, learning target-specific features is pivotal in enhancing detector performance. However, previous methods mostly concentrated on aligning domain-invariant features across domains and neglected integrating the specific features. To tackle this issue, we introduce a novel feature learning method called Joint Feature Differentiation and Interaction (JFDI), which significantly boosts the adaptability of the object detector. We construct a dual-path architecture based on we proposed feature differentiate modules: One path, guided by the source domain data, utilizes multiple discriminators to confuse and align domain-invariant features. The other path, specifically tailored to the target domain, learns its distinctive characteristics based on pseudo-labeled target data. Subsequently, we implement an interactive enhanced mechanism between these paths to ensure stable learning of features and mitigate interference from pseudo-label noise during the iterative optimization. Additionally, we devise a hierarchical pseudo-label fusion module that consolidates more comprehensive and reliable results. In addition, we analyze the generalization error bound of JFDI, which provides a theoretical basis for the effectiveness of JFDI. Extensive empirical evaluations across diverse benchmark scenarios demonstrate that our method is advanced and efficient.

Single-channel speech enhancement primarily relies on deep learning models to recover clean speech signals from noise-contaminated speech. These models establish a mapping relationship between noisy and clean speech. However, considering the sparse distribution characteristics of speech energy across the entire time–frequency spectrogram, constructing the mapping relationship from noisy to clean speech exhibits significant differences in regions where speech energy is concentrated and non-concentrated. Utilizing one deep model to simultaneously address these two distinct regression tasks increases the complexity of the mapping relationships, consequently restricting the model’s performance. To validate our hypothesis, we propose a dual-region speech enhancement model based on voiceprint region segmentation. Specifically, we first train a voiceprint segmentation model to classify noisy speech into two regions. Subsequently, we establish dedicated speech enhancement models for each region, with the dual-region models concurrently constructing mapping relationships for noise-corrupted speech to clean speech in distinct regions. Finally, by merging the results, the complete restored speech can be obtained. Experimental results on public datasets demonstrate that our method achieves competitive speech enhancement performance, outperforming the state-of-the-art. Ablation study results confirm the effectiveness of the proposed approach in enhancing model performance.

Humans have the ability to constantly learn new knowledge. However, for artificial intelligence, trying to continuously learn new knowledge usually results in catastrophic forgetting, the existing regularization-based and dynamic structure-based approaches have shown great potential for alleviating. Nevertheless, these approaches have certain limitations. They usually do not fully consider the problem of incompatible feature embeddings. Instead, they tend to focus only on the features of new or previous classes and fail to comprehensively consider the entire model. Therefore, we propose a two-stage learning paradigm to solve feature embedding incompatibility problems. Specifically, we retain the previous model and freeze all its parameters in the first stage while dynamically expanding a new module to alleviate feature embedding incompatibility questions. In the second stage, a fusion knowledge distillation approach is used to compress the redundant feature dimensions. Moreover, we propose weight pruning and consolidation approaches to improve the efficiency of the model. Our experimental results obtained on the CIFAR-100, ImageNet-100 and ImageNet-1000 benchmark datasets show that the proposed approaches achieve the best performance among all the compared approaches. For example, on the ImageNet-100 dataset, the maximal accuracy improvement is 5.08%. Code is available at https://github.com/ybyangjing/CIL-FCE.

This note aims to reduce the conservatism of stability criteria for neural networks with time-varying delay. To this goal, on the one hand, we construct an augmented Lyapunov–Krasovskii functional (LKF), incorporating some delay-product terms that capture more information about neural states. On the other hand, when dealing with the derivative of the LKF, we introduce several parameter-dependent slack matrices into an affine integral inequality, zero equations, and the $S$-procedure. As a result, more relaxed stability criteria are obtained by employing the so-called Lyapunov–Krasovskii Theorem. Two numerical examples show that the proposed stability criteria are of less conservatism compared with some existing methods.

Image harmonization seeks to transfer the illumination distribution of the background to that of the foreground within a composite image. Existing methods lack the ability of establishing global–local pixel illumination dependencies between foreground and background of composite images, which is indispensable for sharp and color-consistent harmonized image generation. To overcome this challenge, we design a novel Simple Hybrid CNN-Transformer Network (SHT-Net), which is formulated into an efficient symmetrical hierarchical architecture. It is composed of two newly designed light-weight Transformer blocks. Firstly, the scale-aware gated block is designed to capture multi-scale features through different heads and expand the receptive fields, which facilitates to generate images with fine-grained details. Secondly, we introduce a simple parallel attention block, which integrates the window-based self-attention and gated channel attention in parallel, resulting in simultaneously global–local pixel illumination relationship modeling capability. Besides, we propose an efficient simple feed forward network to filter out less informative features and allow the features to contribute to generating photo-realistic harmonized results passing through. Extensive experiments on image harmonization benchmarks indicate that our method achieve promising quantitative and qualitative results. The code and pre-trained models are available at https://github.com/guanguanboy/SHT-Net.

Multi-view multi-label learning (MVML) aims to train a model that can explore the multi-view information of the input sample to obtain its accurate predictions of multiple labels. Unfortunately, a majority of existing MVML methods are based on the assumption of data completeness, making them useless in practical applications with partially missing views or some uncertain labels. Recently, many approaches have been proposed for incomplete data, but few of them can handle the case of both missing views and labels. Moreover, these few existing works commonly ignore potentially valuable information about unknown labels or do not sufficiently explore latent label information. Therefore, in this paper, we propose a label semantic-guided contrastive learning method named LSGC for the dual incomplete multi-view multi-label classification problem. Concretely, LSGC employs deep neural networks to extract high-level features of samples. Inspired by the observation of exploiting label correlations to improve the feature discriminability, we introduce a graph convolutional network to effectively capture label semantics. Furthermore, we introduce a new sample-label contrastive loss to explore the label semantic information and enhance the feature representation learning. For missing labels, we adopt a pseudo-label filling strategy and develop a weighting mechanism to explore the confidently recovered label information. We validate the framework on five standard datasets and the experimental results show that our method achieves superior performance in comparison with the state-of-the-art methods.

Over the past decades, massive Electronic Health Records (EHRs) have been accumulated in Intensive Care Unit (ICU) and many other healthcare scenarios. The rich and comprehensive information recorded presents an exceptional opportunity for patient outcome predictions. Nevertheless, due to the diversity of data modalities, EHRs exhibit a heterogeneous characteristic, raising a difficulty to organically leverage information from various modalities. It is an urgent need to capture the underlying correlations among different modalities. In this paper, we propose a novel framework named Multimodal Fusion Network (MFNet) for ICU patient outcome prediction. First, we incorporate multiple modality-specific encoders to learn different modality representations. Notably, a graph guided encoder is designed to capture underlying global relationships among medical codes, and a text encoder with pre-fine-tuning strategy is adopted to extract appropriate text representations. Second, we propose to pairwise merge multimodal representations with a tailored hierarchical fusion mechanism. The experiments conducted on the eICU-CRD dataset validate that MFNet achieves superior performance on mortality prediction and Length of Stay (LoS) prediction compared with various representative and state-of-the-art baselines. Moreover, comprehensive ablation study demonstrates the effectiveness of each component of MFNet.

Inertial neural networks are proposed via introducing an inertia term into the Hopfield models, which make their dynamic behavior more complex compared to the traditional first-order models. Besides, the aperiodically intermittent quantized control over conventional feedback control has its potential advantages on reducing communication blocking and saving control cost. Based on these facts, we are mainly devoted to exploring of exponential synchronization of quaternion-valued inertial neural networks under aperiodically intermittent quantized control. Firstly, a compact quaternion-valued aperiodically intermittent quantized control protocol is developed, which can mitigate significantly the complexity of theoretical derivation. Subsequently, several concise criteria involving matrix inequalities are formulated through constructing a type of Lyapunov functional and employing a direct analysis approach. The correctness of the obtained results eventually is verified by a typical example.

Unsupervised graph learning techniques have garnered increasing interest among researchers. These methods employ the technique of maximizing mutual information to generate representations of nodes and graphs. We show that these methods are susceptible to backdoor attacks, wherein the adversary can poison a small portion of unlabeled graph data (e.g., node features and graph structure) by introducing triggers into the graph. This tampering disrupts the representations and increases the risk to various downstream applications. Previous backdoor attacks in supervised learning primarily operate directly on the label space and may not be suitable for unlabeled graph data. To tackle this challenge, we introduce GRBA,¹ a gradient-based first-order backdoor attack method. To the best of our knowledge, this constitutes a pioneering endeavor in investigating backdoor attacks within the domain of unsupervised graph learning. The initiation of this method does not necessitate prior knowledge of downstream tasks, as it directly focuses on representations. Furthermore, it is versatile and can be applied to various downstream tasks, including node classification, node clustering and graph classification. We evaluate GRBA on state-of-the-art unsupervised learning models, and the experimental results substantiate the effectiveness and evasiveness of GRBA in both node-level and graph-level tasks.