Local Embedding Research Articles

Understanding the phenomenon of food waste valorisation from the perspective of supply chain actors engaged in it

The repurposing of surplus food and food processing by-products is a key aspect of the shift towards a circular bioeconomy. In the Netherlands, food supply chain actors are already working towards making better use of agri-food surpluses to improve food security and resource efficiency. By analysing the experiences of these actors, this study aims to identify the current state of the sector and highlight barriers and enablers of successful food waste valorisation. Based on a phenomenological analysis of 16 in-depth unstructured interviews with participants from the Netherlands, we empirically demonstrate that supply chain actors involved in food waste valorisation consider local embedding, societal perception, and food safety to be important aspects of their work. They often grapple with legislative and logistical uncertainties and face various ethical dilemmas such as having to pay for products that were once considered waste. Building on these findings, we propose that experienced professionals with a penchant for sustainability can play the role of ‘transition brokers’ as the food system moves towards a circular way of operating. We also posit that private standards and voluntary agreements could be a way for the sector to better manage the various bottlenecks that are characteristic of food waste valorisation endeavours.

Quantum computation of reactions on surfaces using local embedding

Modeling electronic systems is an important application for quantum computers. In the context of materials science, an important open problem is the computational description of chemical reactions on surfaces. In this work, we outline a workflow to model the adsorption and reaction of molecules on surfaces using quantum computing algorithms. We develop and compare two local embedding methods for the systematic determination of active spaces. These methods are automated and based on the physics of molecule-surface interactions and yield systematically improvable active spaces. Furthermore, to reduce the quantum resources required for the simulation of the selected active spaces using quantum algorithms, we introduce a technique for exact and automated circuit simplification. This technique is applicable to a broad class of quantum circuits and critical to enable demonstration on near-term quantum devices. We apply the proposed combination of active-space selection and circuit simplification to the dissociation of water on a magnesium surface using classical simulators and quantum hardware. Our study identifies reactions of molecules on surfaces, in conjunction with the proposed algorithmic workflow, as a promising research direction in the field of quantum computing applied to materials science.

Cross-domain privacy-preserving broad network for fault diagnosis of rotating machinery

The broad learning system-based data processing method has been widely used in the field of industrial intelligent operation and maintenance, and has achieved impressive results. However, data island created by privacy concerns in real industrial scenarios makes centralized deep training extremely difficult. In addition, the fluctuating operating conditions of machines including speed and load significantly affect the distribution of monitoring data representing the operating state of the device. These two problems make it difficult to apply broad learning system to practical application scenarios. Therefore, in view of the above problems, a novel cross-domain privacy-preserving broad network (CDPPBN) for cross-domain fault classification of rotating machinery is presented. Specifically, a cross-domain interactive data encryption protocol feature extractor is designed firstly to ensure the privacy and security of both source and target domain data during feature extraction process. Then, a domain adaptation target classifier is constructed to achieve the fault classification task in the target domain without labels. In this classifier, a maximum mean discrepancy regularization term is constructed to match the mean projections of the source and target domains to adapt the source classifier to the target domain, and a semi-supervised local graph embedding regularization term is constructed to mine the structural features of the target domain data. CDPPBN not only has high performance in terms of data privacy protection, but also makes full use of both labeled source data and unlabeled target data to generate a domain adaptation target classifier. The effectiveness of the presented method is verified through bearing monitoring signal. Experimental analysis shows that the proposed CDPPBN has better data privacy protection capability and data cross-domain diagnosis ability compared with other methods.

Fractionalization paves the way to local projector embeddings of quantum many-body scars

Fractionalization paves the way to local projector embeddings of quantum many-body scars

Modified locally joint sparse marginal embedding and wasserstein generation adversarial network for bearing fault diagnosis

Rolling bearings are essential parts for manufacturing machines. Vast quantities of features are often extracted from measured signals to comprehensively reflect the conditions of bearings, which may cause high dimensionality, information redundancy, and time consumption. In addition, it is extremely difficult, expensive, and time-consuming to collect samples with label information during the bearing fault diagnosis in real-world scenarios. In this study, a novel bearing defect diagnosis method for small sample size is proposed based on modified local joint sparse marginal embedding (MLJSME) and Wasserstein generative adversarial networks (WGANs). MLJSME can effectively extract intrinsic sparse discriminant features of high-dimensional dataset by preserving both global and local structures. Graph embedding and Gaussian kernel function are adopted to preserve the locality structure of dataset. The global structure and discriminate information are preserved by maximum margin criterion which can also avoid small sample-size problem. Moreover, joint sparsity is applied to preserve the sparse property and improve the robustness to noise and outliers. An abundance of artificial samples can be obtained with WGAN and a few labeled samples. Firstly, a high-dimensional feature dataset consisting of time-domain and frequency-domain features is extracted from original vibration signals, then MLJSME is utilized to extract sensitive low-dimensional features, and a small number of low-dimensional features are fed into WGAN to generate a large number of artificial samples that used to train the classifier, and the bearing fault types can be finally identified. The effectiveness and feasibility of the proposed method is validated by analyzing the different experimental cases.

A global–local context embedding learning based sequence-free framework for state of health estimation of lithium-ion battery

Accurate estimation of the state of health (SOH) of lithium-ion batteries holds significant importance in guaranteeing the stable and secure functioning of electric vehicles. However, existing neural network-based methods suffer from limitations in capturing long-term serial relationships and extracting degenerate features. In light of these challenges, we propose a novel sequence-free framework for performing the SOH estimation task. Technically, a global–local context embedding module is proposed to learn both global- and local-range information context by two convolutional streams with different depths. With this module, a discriminatory feature learning can be guided. By integrating it into the convolution neural network, a novel time series prediction network, called improved convolution neural network (ICNN) is presented, which can effectively establish the mapping relationship between battery charging/discharging curves and battery SOH. Through rigorous experimentation on the CACLE dataset and NASA dataset, we demonstrate the efficacy of our proposed method, achieving mean absolute errors (MAEs) of 0.54% and 1.20% respectively. Our findings highlight the superiority of the proposed method compared to commonly used neural network methods in the domain of battery SOH estimation.

Experimental study on fracture damage and seismic performance of loose through-tenon joints in ancient timber structures

A looseness is typical damage of mortise-tenon joints in ancient timber structures. Brittle fractures of loose joints are likely to occur, leading to joint failure during earthquakes. This study conducts pseudo-static tests on three groups of full-scaled through-tenon joints to examine the fracture damage and seismic behavior of the loose mortise-tenon joints and determines fracture damage modes of through-tenon joints with different looseness. It uses the acoustic emission method to measure the ring count on the tenon surface, along with each joint's critical load and critical energy release rate and then analyzes the fracture damage performance of different tenon joints. In addition, this research investigates the hysteretic and skeleton curve, stiffness degradation curve, and tenon pulling quantity of the joint under various fracture damage states. The results showed that the fracture failure mode of the through-tenon joint is a fracture in the variable cross-section and along the local tenon neck. During the loading process, the critical energy release rate initially increases and then decreases as the looseness grows. In addition, with the increase of crack mouth opening displacement (CMOD) and crack length, the bending moment of the joint firstly increases and then decreases. When the variable cross-section breaks and the tenon neck partially cracks, the joints' bending moment and stiffness reduce with increasing looseness. With the increase in local embedment compression in the lower side of the beam end and the mortise, the amount of tenon pull-out of loose joints decreases under positive and negative loading.

DANet: Temporal Action Localization with Double Attention

Temporal action localization (TAL) aims to predict action instance categories in videos and identify their start and end times. However, existing Transformer-based backbones focus only on global or local features, resulting in the loss of information. In addition, both global and local self-attention mechanisms tend to average embeddings, thereby reducing the preservation of critical features. To solve these two problems better, we propose two kinds of attention mechanisms, namely multi-headed local self-attention (MLSA) and max-average pooling attention (MA) to extract simultaneously local and global features. In MA, max-pooling is used to select the most critical information from local clip embeddings instead of averaging embeddings, and average-pooling is used to aggregate global features. We use MLSA for modeling local temporal context. In addition, to enhance collaboration between MA and MLSA, we propose the double attention block (DABlock), comprising MA and MLSA. Finally, we propose the final network double attention network (DANet), composed of DABlocks and other advanced blocks. To evaluate DANet’s performance, we conduct extensive experiments for the TAL task. Experimental results demonstrate that DANet outperforms the other state-of-the-art models on all datasets. Finally, ablation studies demonstrate the effectiveness of the proposed MLSA and MA. Compared with structures using backbone with convolution and global Transformer, DABlock consisting of MLSA and MA has a superior performance, achieving an 8% and 0.5% improvement on overall average mAP, respectively.

Weakly supervised point cloud segmentation via deep morphological semantic information embedding

AbstractSegmenting the semantic regions of point clouds is a crucial step for intelligent agents to understand 3D scenes. Weakly supervised point cloud segmentation is highly desirable because entirely labelling point clouds is highly time‐consuming and costly. For the low‐costing labelling of 3D point clouds, the scene‐level label is one of the most effortless label strategies. However, due to the limitation of classifier discriminative capability and the orderless and structurless nature of the point cloud data, existing scene‐level method is hard to transfer the semantic information, which usually leads to the under‐activated or over‐activated issues. To this end, a local semantic embedding network is introduced to learn local structural patterns and semantic propagation. Specifically, the proposed network contains graph convolution‐based dilation and erosion embedding modules to implement ‘inside‐out’ and ‘outside‐in’ semantic information dissemination pathways. Therefore, the proposed weakly supervised learning framework could achieve the mutual propagation of semantic information in the foreground and background. Comprehensive experiments on the widely used ScanNet benchmark demonstrate the superior capacity of the proposed approach when compared to the current alternatives and baseline models.

Beyond independent error assumptions in large GNN atomistic models.

The calculation of relative energy difference has significant practical applications, such as determining adsorption energy, screening for optimal catalysts with volcano plots, and calculating reaction energies. Although Density Functional Theory (DFT) is effective in calculating relative energies through systematic error cancellation, the accuracy of Graph Neural Networks (GNNs) in this regard remains uncertain. To address this, we analyzed ∼483 × 106 pairs of energy differences predicted by DFT and GNNs using the Open Catalyst 2020-Dense dataset. Our analysis revealed that GNNs exhibit a correlated error that can be reduced through subtraction, challenging the assumption of independent errors in GNN predictions and leading to more precise energy difference predictions. To assess the magnitude of error cancellation in chemically similar pairs, we introduced a new metric, the subgroup error cancellation ratio. Our findings suggest that state-of-the-art GNN models can achieve error reduction of up to 77% in these subgroups, which is comparable to the error cancellation observed with DFT. This significant error cancellation allows GNNs to achieve higher accuracy than individual energy predictions and distinguish subtle energy differences. We propose the marginal correct signratio as a metric to evaluate this performance. Additionally, our results show that the similarity in local embeddings is related to the magnitude of error cancellation, indicating the need for a proper training method that can augment the embedding similarity for chemically similar adsorbate-catalyst systems.

MSaD-Net: A Mix Self-Attention Networks for 3D Point Cloud Denoising

In the process of acquiring 3D point cloud data, due to environmental interference or unstable scanning equipment, the acquired data often have noisy points. Recently, with the development of neural networks for point clouds, great progress has been made in deep learning-based point cloud denoising. However, most of the existing methods adopt a pointnet-like structure to predict point offsets. Simple pooling operation loses much important information, such as local neighborhood information, and global information. The loss of information makes the algorithm ineffective when dealing with some complex cases. In order to solve the above problems, we propose a self-attention-based point cloud denoising network architecture, through the Transformer structure, to establish long-range dependencies of the points. In addition, we propose a local information embedding module to fast select meaningful points and serve as the input of the Transformer. We also consider the correlation between channels of point cloud features and further introduce a channel attention module. Extensive experiments prove that our method outperforms existing methods, and maintains a high running speed.

Using Hybrid HMM/DNN Embedding Extractor Models in Computational Paralinguistic Tasks.

The field of computational paralinguistics emerged from automatic speech processing, and it covers a wide range of tasks involving different phenomena present in human speech. It focuses on the non-verbal content of human speech, including tasks such as spoken emotion recognition, conflict intensity estimation and sleepiness detection from speech, showing straightforward application possibilities for remote monitoring with acoustic sensors. The two main technical issues present in computational paralinguistics are (1) handling varying-length utterances with traditional classifiers and (2) training models on relatively small corpora. In this study, we present a method that combines automatic speech recognition and paralinguistic approaches, which is able to handle both of these technical issues. That is, we trained a HMM/DNN hybrid acoustic model on a general ASR corpus, which was then used as a source of embeddings employed as features for several paralinguistic tasks. To convert the local embeddings into utterance-level features, we experimented with five different aggregation methods, namely mean, standard deviation, skewness, kurtosis and the ratio of non-zero activations. Our results show that the proposed feature extraction technique consistently outperforms the widely used x-vector method used as the baseline, independently of the actual paralinguistic task investigated. Furthermore, the aggregation techniques could be combined effectively as well, leading to further improvements depending on the task and the layer of the neural network serving as the source of the local embeddings. Overall, based on our experimental results, the proposed method can be considered as a competitive and resource-efficient approach for a wide range of computational paralinguistic tasks.

Few-shot IoT attack detection based on SSDSAE and adaptive loss weighted meta residual network

The Internet of Things (IoT) is an open and comprehensive network of smart objects. Unfortunately, it is also becoming increasingly vulnerable to security attacks during the increasing popularity of the IoT. It can lead traditional antivirus software to be less likely to prevent this threat. Therefore, it is necessary to design a model for the IoT attack detection. Current detection models are trained using massive big data samples. However, the distribution of traffic samples is few in specific scenarios. Also, existing models are also susceptible to noise interference in IoT environments, lowering detection efficiency and accuracy. In this work, we propose a few-shot IoT attack detection approach using a semi-supervised deep sparse autoencoder (SSDSAE) and an adaptive loss weighted meta residual network (ALWM-ResNet). First, an SSDSAE feature extraction model for local graph embedding is designed using local and non-local graph embedding constraints. Then, we design an ALWM-ResNet model to achieve IoT attack detection with few-shot samples under noise labels. A weighted function map is established using a weighted network and a meta-model, and weights are adaptively learned from the noise labels. Finally, we validate our approach using four IoT datasets. Several experimental results demonstrate the superior performance of our approach in IoT attack detection under few-shot samples.

TransPose Re-ID: transformers for pose invariant person Re-identification

ABSTRACT Person re-identification (Re-ID) is a computer vision task that involves recognizing and tracking individuals across multiple non-overlapping cameras or over time within the same camera view. It is particularly important in surveillance systems, where it can help in identifying potential threats or tracking suspects. Convolutional neural networks (CNNs) have been used to extract invariant person representation for this challenging task. However, CNNs do not consider global dependencies in their initial layers, causing some vital information to be lost during the convolution process. The development of vision-based transformers has opened up new research avenues for person re-identification. This work proposes a purely transformer-based solution, called TansPose Re-ID, that learns pose-invariant person representations. The proposed system uses a vision transformer baseline and enhances its architecture by introducing multiple streams to learn global and local dependencies as well as pose invariance in person images. The architecture includes a Global Self-Attention Module (GSM) and a Local Self-Attention Module (LSM) that jointly learn global and local patch-based person embeddings. The LSM is further improved by stochastically grouping local patches and aligning them. Additionally, an attention feature learning module (AFLM) is introduced in the LSM to handle pose and viewpoint variations. The proposed method is evaluated on two public Re-ID benchmarks, Market1501 and DukeMTMC-ReID, and demonstrates superior performance compared to existing transformer baselines.

Dual-branch vision transformer for blind image quality assessment

Blind image quality assessment (BIQA) has always been a challenging problem due to the absence of reference images. In this paper, we propose a novel dual-branch vision transformer for BIQA, which simultaneously considers both local distortions and global semantic information. It first extracts dual-scale features from the backbone network, and then each scale feature is fed into one of the transformer encoder branches as a local feature embedding to consider the scale-variant local distortions. Each transformer branch obtains the context of global image distortion as well as the local distortion by adopting content-aware embedding. Finally, the outputs of the dual-branch vision transformer are combined by using multiple feed-forward blocks to predict the image quality scores effectively. Experimental results demonstrate that the proposed BIQA method outperforms the conventional methods on the six public BIQA datasets.

Robust Discriminant Subspace Clustering With Adaptive Local Structure Embedding.

Unsupervised dimension reduction and clustering are frequently used as two separate steps to conduct clustering tasks in subspace. However, the two-step clustering methods may not necessarily reflect the cluster structure in the subspace. In addition, the existing subspace clustering methods do not consider the relationship between the low-dimensional representation and local structure in the input space. To address the above issues, we propose a robust discriminant subspace (RDS) clustering model with adaptive local structure embedding. Specifically, unlike the existing methods which incorporate dimension reduction and clustering via regularizer, thereby introducing extra parameters, RDS first integrates them into a unified matrix factorization (MF) model through theoretical proof. Furthermore, a similarity graph is constructed to learn the local structure. A constraint is imposed on the graph to guarantee that it has the same connected components with low-dimensional representation. In this spirit, the similarity graph serves as a tradeoff that adaptively balances the learning process between the low-dimensional space and the original space. Finally, RDS adopts the l2,1 -norm to measure the residual error, which enhances the robustness to noise. Using the property of the l2,1 -norm, RDS can be optimized efficiently without introducing more penalty terms. Experimental results on real-world benchmark datasets show that RDS can provide more interpretable clustering results and also outperform other state-of-the-art alternatives.

Feature preserving 3D mesh denoising with a Dense Local Graph Neural Network

Graph neural networks (GNNs) are ideally suited for mesh denoising. However, existing solutions such as those based on graph convolutional networks (GCNs) are built for a fixed graph thus making them not naturally generalizable to unseen meshes. Furthermore, their graph Laplacian based global node embedding algorithms can cause excessive smoothing while achieving feature preserving mesh denoising requires a GNN to possess local processing capability. This paper presents a mesh denoising method via a new Dense lOcal Graph neural NETwork (DOGNET). DOGNET implements a local node embedding algorithm that generates node embeddings through aggregating information from a node’s connected local neighbours which automatically make DOGNET inductive as well as effective for feature preserving mesh denoising. We present extensive experimental results to demonstrate quantitatively and qualitatively that DOGNET is superior to SOTA meshing denoising techniques.

Contrastive Graph Representations for Logical Formulas Embedding

Currently, the non-transparent computing process of deep learning has become a significant reason hindering its further development. The Neural-Symbolic (NS) system formed by integrating logic rules into neural networks has attracted increasing attention owing to its direct interpretability. Embedding symbolic logical formulas into a low-dimensional continuous space provides an effective way for the NS system. However, current studies are all constrained by the modeling ability for its syntactic structure and fail to preserve the intrinsic semantics in embeddings, which causes poor performance on downstream reasoning tasks. To this end, this paper proposes a novel method of <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Con</b> trastive <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">G</b> raph <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">R</b> epresentations (ConGR) for logical formulas embedding. First, to improve the modeling ability for the syntactic structure, ConGR introduces a densely connected graph convolutional network (GCN) with an attention mechanism to process syntax parsing graphs of formulas. In this way, discriminative local and global embeddings of formulas are obtained at the syntax level. Second, the contrastive instances (positive or negative) for each anchor formula are generated by the transformation under the guidance of logical properties. To preserve semantic information, two types of contrast, global-local and global-global, are carried out to refine formula embeddings. Extensive experiments demonstrate that ConGR obtains superior performance against state-of-the-art baselines on entailment checking and premise selection datasets.

Hierarchical federated learning with local model embedding

Federated learning can synergize the local model training with private data samples from geo-distributed users. Nevertheless, the unification process of a comprehensive global model through periodical parameter sharing can be time-consuming at a high cost. On the one hand, the data samples are collected from users with diverse preferences, and the data distribution can be non-independent and identically distributed (non-IID). On the other hand, the consequent model trained locally needs to communicate with a remote parameter server periodically for parameter synchronization, which leads to overwhelming communication and synchronization overhead given heterogeneous device capacities and network conditions of end-users. Generally, a hierarchical system design with a clustered group is ideal for accommodating diversity. Actually, it is still challenging to maintain the relationship of the local model training without knowing the data samples in advance for privacy concerns. Therefore, we present hierarchical federated model embedding to formulate the relationship between local data distributions. Initially, the local models are embedded through the global shared dataset to obtain feature latent representation vectors. The cloud server groups the clients according to the vectors, making clients with similar data distribution train collaboratively in a same group. Then, these vectors are used to train the predictor on the cloud server, which is utilized for efficient group assignment when new clients join the system. Compared with the baseline, the accuracy of the group model can be improved by 1.22%∼5.63% and that of the global model can be improved by 3.97%∼14.25% on different datasets.

MFPCDR: A Meta-Learning-Based Model for Federated Personalized Cross-Domain Recommendation

Cross-domain recommendation systems frequently require the use of rich source domain information to improve recommendations in the target domain, thereby resolving the data sparsity and cold-start problems, whereas the majority of existing approaches frequently require the centralized storage of user data, which poses a substantial risk of privacy breaches. Compared to traditional recommendation systems with centralized data, federated recommendation systems with multiple clients trained collaboratively have significant privacy benefits in terms of user data. While users’ interests are often personalized, meta-learning can be used to learn users’ personalized preferences, and personalized preferences can help models make recommendations in cold-start scenarios. We use meta-learning to learn the personalized preferences of cold-start users. Therefore, we offer a unique meta-learning-based federated personalized cross-domain recommendation model that discovers the personalized preferences for cold-start users via a server-side meta-recommendation module. To avoid compromising user privacy, an attention mechanism is used on each client to find transferable features that contribute to knowledge transfer while obtaining embeddings of users and items; each client then uploads the weights to the server. The server accumulates weights and delivers them to clients for update. Compared to traditional recommendation system models, our model can effectively protect user privacy while solving the user cold-start problem, as we use an attention mechanism in the local embedding module to mine the source domain for transferable features that contribute to knowledge transfer. Extensive trials on real-world datasets have demonstrated that our technique effectively guarantees speed while protecting user privacy.