Neural Networks最新文献

Improperly-exposed images often have unsatisfactory visual characteristics like inadequate illumination, low contrast, and the loss of small structures and details. The mapping relationship from an improperly-exposed condition to a well-exposed one may vary significantly due to the presence of multiple exposure conditions. Consequently, the enhancement methods that do not pay specific attention to this issue tend to yield inconsistent results when applied to the same scene under different exposure conditions. In order to obtain consistent enhancement results for various exposures while restoring rich details, we propose an illumination-aware divide-and-conquer network (IDNet). Specifically, to address the challenge of directly learning a sophisticated nonlinear mapping from an improperly-exposed condition to a well-exposed one, we utilize the discrete wavelet transform (DWT) to decompose the image into the low-frequency (LF) component, which primarily captures brightness and contrast, and the high-frequency (HF) components that depict fine-scale structures. To mitigate the inconsistency in correction across various exposures, we extract a conditional feature from the input that represents illumination-related global information. This feature is then utilized to modulate the dynamic convolution weights, enabling precise correction of the LF component. Furthermore, as the co-located positions of LF and HF components are highly correlated, we create a mask to distill useful knowledge from the corrected LF component, and integrate it into the HF component to support the restoration of fine-scale details. Extensive experimental results demonstrate that the proposed IDNet is superior to several state-of-the-art enhancement methods on two datasets with multiple exposures.

Adversarial attacks dramatically change the output of an otherwise accurate learning system using a seemingly inconsequential modification to a piece of input data. Paradoxically, empirical evidence indicates that even systems which are robust to large random perturbations of the input data remain susceptible to small, easily constructed, adversarial perturbations of their inputs. Here, we show that this may be seen as a fundamental feature of classifiers working with high dimensional input data. We introduce a simple generic and generalisable framework for which key behaviours observed in practical systems arise with high probability—notably the simultaneous susceptibility of the (otherwise accurate) model to easily constructed adversarial attacks, and robustness to random perturbations of the input data. We confirm that the same phenomena are directly observed in practical neural networks trained on standard image classification problems, where even large additive random noise fails to trigger the adversarial instability of the network. A surprising takeaway is that even small margins separating a classifier’s decision surface from training and testing data can hide adversarial susceptibility from being detected using randomly sampled perturbations. Counter-intuitively, using additive noise during training or testing is therefore inefficient for eradicating or detecting adversarial examples, and more demanding adversarial training is required.

Local feature extraction plays a crucial role in numerous critical visual tasks. However, there remains room for improvement in both descriptors and keypoints, particularly regarding the discriminative power of descriptors and the localization precision of keypoints. To address these challenges, this study introduces a novel local feature extraction pipeline named OSDFeat (Object and Spatial Discrimination Feature). OSDFeat employs a decoupling strategy, training descriptor and detection networks independently. Inspired by semantic correspondence, we propose an Object and Spatial Discrimination ResUNet (OSD-ResUNet). OSD-ResUNet captures features from the feature map that differentiate object appearance and spatial context, thus enhancing descriptor performance. To further improve the discriminative capability of descriptors, we propose a Discrimination Information Retained Normalization module (DIRN). DIRN complementarily integrates spatial-wise normalization and channel-wise normalization, yielding descriptors that are more distinguishable and informative. In the detection network, we propose a Cross Saliency Pooling module (CSP). CSP employs a cross-shaped kernel to aggregate long-range context in both vertical and horizontal dimensions. By enhancing the saliency of keypoints, CSP enables the detection network to effectively utilize descriptor information and achieve more precise localization of keypoints. Compared to the previous best local feature extraction methods, OSDFeat achieves Mean Matching Accuracy of 79.4% in local feature matching task, improving by 1.9% and achieving state-of-the-art results. Additionally, OSDFeat achieves competitive results in Visual Localization and 3D Reconstruction. The results of this study indicate that object and spatial discrimination can improve the accuracy and robustness of local feature, even in challenging environments. The code is available at https://github.com/pandaandyy/OSDFeat.

With the rapid advent and abundance of remote sensing data in different modalities, cross-modal retrieval tasks have gained importance in the research community. Cross-modal retrieval belongs to the research paradigm in which the query is of one modality and the retrieved output is of the other modality. In this paper, the remote sensing (RS) data modalities considered are the earth observation optical data (aerial photos) and the corresponding hand-drawn sketches. The main challenge of the cross-modal retrieval research objective for optical remote sensing images and the corresponding sketches is the distribution gap between the shared embedding space of the modalities. Prior attempts to resolve this issue have not yielded satisfactory outcomes regarding accurately retrieving cross-modal sketch-image RS data. The state-of-the-art architectures used conventional convolutional architectures, which focused on local pixel-wise information about the modalities to be retrieved. This limits the interaction between the sketch texture and the corresponding image, making these models susceptible to overfitting datasets with particular scenarios. To circumvent this limitation, we suggest establishing multi-modal correspondence using a novel architecture of the combined self and cross-attention algorithms, SPCA-Net to minimize the modality gap by employing attention mechanisms for the query and other modalities. Efficient cross-modal retrieval is achieved through the suggested attention architecture, which empirically emphasizes the global information of the relevant query modality and bridges the domain gap through a unique pairwise cross-attention network. In addition to the novel architecture, this paper introduces a unique loss function, label-specific supervised contrastive loss, tailored to the intricacies of the task and to enhance the discriminative power of the learned embeddings. Extensive evaluations are conducted on two sketch-image remote sensing datasets, Earth-on-Canvas and RSketch. Under the same experimental conditions, the performance metrics of our proposed model beat the state-of-the-art architectures by significant margins of 16.7%, 18.9%, 33.7%, and 40.9% correspondingly.

Estimating intracranial current sources underlying the electromagnetic signals observed from extracranial sensors is a perennial challenge in non-invasive neuroimaging. Established solutions to this inverse problem treat time samples independently without considering the temporal dynamics of event-related brain processes.

This paper describes current source estimation from simultaneously recorded magneto- and electro-encephalography (MEEG) using a recurrent neural network (RNN) that learns sequential relationships from neural data. The RNN was trained in two phases: (1) pre-training and (2) transfer learning with L1 regularization applied to the source estimation layer. Performance of using scaled labels derived from MEEG, magnetoencephalography (MEG), or electroencephalography (EEG) were compared, as were results from volumetric source space with free dipole orientation and surface source space with fixed dipole orientation. Exact low-resolution electromagnetic tomography (eLORETA) and mixed-norm L1/L2 (MxNE) source estimation methods were also applied to these data for comparison with the RNN method.

The RNN approach outperformed other methods in terms of output signal-to-noise ratio, correlation and mean-squared error metrics evaluated against reference event-related field (ERF) and event-related potential (ERP) waveforms. Using MEEG labels with fixed-orientation surface sources produced the most consistent estimates.

To estimate sources of ERF and ERP waveforms, the RNN generates temporal dynamics within its internal computational units, driven by sequential structure in neural data used as training labels. It thus provides a data-driven model of computational transformations from psychophysiological events into corresponding event-related neural signals, which is unique among MEEG source reconstruction solutions.

Real-world data is typically distributed on low-dimensional manifolds embedded in high-dimensional Euclidean spaces. Accurately extracting spatial distribution features on general manifolds that reflect the intrinsic characteristics of data is crucial for effective feature representation. Therefore, we propose a generalized geodesic basis function neural network (G${}^2$BFNN) architecture. The generalized geodesic basis functions (G${}^2$BF) are defined based on generalized geodesic distances. The generalized geodesic distance metric (G${}^2$DM) is obtained by learning the manifold structure. To implement this architecture, a specific G${}^2$BFNN, named discriminative local preserving projection-based G${}^2$BFNN (DLPP-G${}^2$BFNN) is proposed. DLPP-G${}^2$BFNN mainly contains two modules, namely the manifold structure learning module (MSLM) and the network mapping module (NMM). In the MSLM module, a supervised adjacency graph matrix is constructed to constrain the learning of the manifold structure. This enables the learned features in the embedding subspace to maintain the manifold structure while enhancing the discriminability. The features and G${}^2$DM learned in the MSLM are fed into the NMM. Through the G${}^2$BF in the NMM, the spatial distribution features on manifold are obtained. Finally, the output of the network is obtained through the fully connected layer. Compared with the local response neural network based on Euclidean distance, the proposed network can reveal more essential spatial structure characteristics of the data. Meanwhile, the proposed G${}^2$BFNN is a generalized network architecture that can be combined with any manifold learning method, showcasing high scalability. The experimental results demonstrate that the proposed DLPP-G${}^2$BFNN outperforms existing methods by utilizing fewer kernels while achieving higher recognition performance.

This study explores the bipartite secure synchronization problem of coupled quaternion-valued neural networks (QVNNs), in which variable sampled communications and random deception attacks are considered. Firstly, by employing the signed graph theory, the mathematical model of coupled QVNNs with structurally-balanced cooperative–competitive interactions is established. Secondly, by adopting non-decomposition method and constructing a suitable unitary Lyapunov functional, the bipartite secure synchronization (BSS) criteria for coupled QVNNs are obtained in the form of quaternion-valued LMIs. It is essential to mention that the structurally-balanced topology is relatively strong, hence, the coupled QVNNs with structurally-unbalanced graph are further studied. The structurally-unbalanced graph is treated as an interruption of the structurally-balanced graph, the bipartite secure quasi-synchronization (BSQS) criteria for coupled QVNNs with structurally-unbalanced graph are derived. Finally, two simulations are given to illustrate the feasibility of the suggested BSS and BSQS approaches.

In the pursuit of potential treatments for neurological disorders and the alleviation of patient suffering, deep brain stimulation (DBS) has been utilized to intervene or investigate pathological neural activities. To explore the exact mechanism of how DBS works, a memristive two-neuron network considering DBS is newly proposed in this work. This network is implemented by coupling two-dimensional Morris–Lecar neuron models and using a memristor synaptic synapse to mimic synaptic plasticity. The complex bursting activities and dynamical effects are revealed numerically through dynamical analysis. By examining the synchronous behavior, the desynchronization mechanism of the memristor synapse is uncovered. The study demonstrates that synaptic connections lead to the appearance of time-lagged or asynchrony in completely synchronized firing activities. Additionally, the memristive two-neuron network is implemented in hardware based on FPGA, and experimental results confirm the abundant neuronal electrical activities and chaotic dynamical behaviors. This work offers insights into the potential mechanisms of DBS intervention in neural networks.

In this work, we propose an end-to-end adaptive sampling framework based on deep neural networks and the moving mesh method (MMPDE-Net), which can adaptively generate new sampling points by solving the moving mesh PDE. This model focuses on improving the quality of sampling points generation. Moreover, we develop an iterative algorithm based on MMPDE-Net, which makes sampling points distribute more precisely and controllably. Since MMPDE-Net is independent of the deep learning solver, we combine it with physics-informed neural networks (PINN) to propose moving sampling PINN (MS-PINN) and show the error estimate of our method under some assumptions. Finally, we demonstrate the performance improvement of MS-PINN compared to PINN through numerical experiments of four typical examples, which numerically verify the effectiveness of our method.