Average Accuracy Improvement Research Articles

Identifying rice field weeds from unmanned aerial vehicle remote sensing imagery using deep learning

BackgroundRice field weed object detection can provide key information on weed species and locations for precise spraying, which is of great significance in actual agricultural production. However, facing the complex and changing real farm environments, traditional object detection methods still have difficulties in identifying small-sized, occluded and densely distributed weed instances. To address these problems, this paper proposes a multi-scale feature enhanced DETR network, named RMS-DETR. By adding multi-scale feature extraction branches on top of DETR, this model fully utilizes the information from different semantic feature layers to improve recognition capability for rice field weeds in real-world scenarios.MethodsIntroducing multi-scale feature layers on the basis of the DETR model, we conduct a differentiated design for different semantic feature layers. The high-level semantic feature layer adopts Transformer structure to extract contextual information between barnyard grass and rice plants. The low-level semantic feature layer uses CNN structure to extract local detail features of barnyard grass. Introducing multi-scale feature layers inevitably leads to increased model computation, thus lowering model inference speed. Therefore, we employ a new type of Pconv (Partial convolution) to replace traditional standard convolutions in the model.ResultsCompared to the original DETR model, our proposed RMS-DETR model achieved an average recognition accuracy improvement of 3.6% and 4.4% on our constructed rice field weeds dataset and the DOTA public dataset, respectively. The average recognition accuracies reached 0.792 and 0.851, respectively. The RMS-DETR model size is 40.8 M with inference time of 0.0081 s. Compared with three classical DETR models (Deformable DETR, Anchor DETR and DAB-DETR), the RMS-DETR model respectively improved average precision by 2.1%, 4.9% and 2.4%.DiscussionThis model is capable of accurately identifying rice field weeds in complex real-world scenarios, thus providing key technical support for precision spraying and management of variable-rate spraying systems.

Design of RISC-V out-of-order processor based on segmented exclusive or Gshare branch prediction

To address the balanced requirements of performance, power consumption, and area in embedded systems, this paper introduces a 32-bit out-of-order processor based on the RISC-V instruction set, and the processor supports interrupt handling. This processor is designed to support the RV32IMC instruction subset and utilizes a four-stage pipeline structure featuring sequential instruction fetching, out-of-order execution, and out-of-order write-back. The main contributions are as follows:1) The segmented exclusive or G-share branch prediction scheme for embedded processors is proposed, which provides high branch prediction accuracy when the capacity of pattern history table (PHT) is small. 2) The work undertaken by hardware and software during interrupt response is reasonably divided, which allows for fast interrupt response with minimal resource consumption. Furthermore, the interrupt response time in vectored mode is reduced by simultaneously stacking the control and status registers (CSRs) and obtaining the interrupt service routine (ISR) entry address. Executing branch prediction-related programs within an identical environment, the processor detailed in this paper demonstrates an average prediction accuracy improvement of 1.2 % over G-share and 0.6 % over bi-mode branch prediction when employing the segmented exclusive or G-share branch prediction scheme. Building on this foundation, the on-chip debugging system and memory management unit have been implemented. The processor reaches 1.389 Dhrystone/MHz and 2.802 Coremark/MHz.

You only compress once: Towards effective and elastic BERT compression via exploit–explore stochastic nature gradient

Despite superior performance on various natural language processing tasks, pre-trained models such as BERT are challenged by deploying on resource-constraint devices. Most existing model compression approaches require re-compression or fine-tuning across diverse constraints to accommodate various hardware deployments. This practically limits the further application of model compression. Moreover, the ineffective training and searching process of existing elastic compression paradigms (Wang et al., 2020; Cai et al., 2020) prevents the direct migration to BERT compression. Motivated by the necessity of efficient inference across various constraints on BERT, we propose a novel approach, YOCO-BERT, to achieve compress once and deploy everywhere. Specifically, we first construct a huge search space with 1013 architectures, which covers nearly all configurations in BERT model. Then, we propose a novel stochastic nature gradient optimization method to guide the generation of optimal candidate architecture which could keep a balanced trade-off between explorations and exploitation. When a certain resource constraint is given, a lightweight distribution optimization approach is utilized to obtain the optimal network for target deployment without fine-tuning. Compared with state-of-the-art algorithms, YOCO-BERT provides more compact models, yet achieving 2.1%–4.5% average accuracy improvement on the GLUE benchmark. Besides, YOCO-BERT is also more effective, e.g., the training complexity is O(1) for N different devices. Codes available https://github.com/MAC-AutoML/YOCO-BERT.

Self-Training-Transductive-Learning Broad Learning System (STTL-BLS): A model for effective and efficient image classification

A novel model called Self-Training-Transductive-Learning Broad Learning System (STTL-BLS) is proposed for image classification. The model consists of two key blocks: Feature Block (FB) and Enhancement Block (EB). The FB utilizes the Proportion of Large Values Attention (PLVA) technique and an Encoder for feature extraction. Multiple FBs are cascaded in the model to learn discriminative features. The Enhancement Block (EB) enhances feature learning and prevents under-fitting on complex datasets. Additionally, an architecture that combines characteristics of Broad Learning System (BLS) and gradient descent is designed for STTL-BLS, enabling the model to leverage the advantages of both BLS and Convolutional Neural Networks (CNNs). Moreover, a training algorithm (STTL) that combines self-training and transductive learning is presented for the model to improve its generalization ability. Experimental results demonstrate that the accuracy of the proposed model surpasses all compared BLS variants and performs comparably or even superior to deep networks: on small-scale datasets, STTL-BLS has an average accuracy improvement of 14.82 percentage points compared to other models; on large-scale datasets, 12.95 percentage points. Notably, the proposed model exhibits low time complexity, particularly with the shortest testing time on the small-scale datasets among all compared models: it has an average testing time of 46.4 s less than other models. It proves to be an additional valuable solution for image classification tasks on both small- and large-scale datasets. The source code for this paper can be accessed at https://github.com/threedteam/sttl_bls.

Data-driven topology optimization using a multitask conditional variational autoencoder with persistent homology

Topology optimization is crucial for the mechanical design of vehicles and aircraft, allowing changes in the shape of structures and the placement of features. Recent advances have integrated deep generative models, particularly convolutional neural networks, to streamline this process.to streamline this process. However, these models struggle to preserve subtle structural features. To overcome these limitations, this study introduced a generative model adept at identifying the topological features inherent in real shapes, such as connectivity and holes, to enhance the effectiveness of topology optimization. A conditional variational autoencoder (CVAE) was employed to predict both the shape and compliance simultaneously. This model, CVAE with persistent homology, generates optimal material distributions by considering topological properties. The learning process introduced a term that minimizes the difference in topological features between true and reconstructed shapes. The proposed model can generate optimal material distributions by considering topological properties, eliminating the need for iterative calculations. This approach was validated using two numerical examples. The accuracy of the generated material distributions was compared with conventional methods using the mean-squared error. An average improvement in accuracy of approximately 36.85% was observed across the two results. This confirms that shapes considering compliance and connectivity can be accurately predicted.

Improving the accuracy of genomic prediction in dairy cattle using the biologically annotated neural networks framework

BackgroundBiologically annotated neural networks (BANNs) are feedforward Bayesian neural network models that utilize partially connected architectures based on SNP-set annotations. As an interpretable neural network, BANNs model SNP and SNP-set effects in their input and hidden layers, respectively. Furthermore, the weights and connections of the network are regarded as random variables with prior distributions reflecting the manifestation of genetic effects at various genomic scales. However, its application in genomic prediction has yet to be explored.ResultsThis study extended the BANNs framework to the area of genomic selection and explored the optimal SNP-set partitioning strategies by using dairy cattle datasets. The SNP-sets were partitioned based on two strategies–gene annotations and 100 kb windows, denoted as BANN_gene and BANN_100kb, respectively. The BANNs model was compared with GBLUP, random forest (RF), BayesB and BayesCπ through five replicates of five-fold cross-validation using genotypic and phenotypic data on milk production traits, type traits, and one health trait of 6,558, 6,210 and 5,962 Chinese Holsteins, respectively. Results showed that the BANNs framework achieves higher genomic prediction accuracy compared to GBLUP, RF and Bayesian methods. Specifically, the BANN_100kb demonstrated superior accuracy and the BANN_gene exhibited generally suboptimal accuracy compared to GBLUP, RF, BayesB and BayesCπ across all traits. The average accuracy improvements of BANN_100kb over GBLUP, RF, BayesB and BayesCπ were 4.86%, 3.95%, 3.84% and 1.92%, and the accuracy of BANN_gene was improved by 3.75%, 2.86%, 2.73% and 0.85% compared to GBLUP, RF, BayesB and BayesCπ, respectively across all seven traits. Meanwhile, both BANN_100kb and BANN_gene yielded lower overall mean square error values than GBLUP, RF and Bayesian methods.ConclusionOur findings demonstrated that the BANNs framework performed better than traditional genomic prediction methods in our tested scenarios, and might serve as a promising alternative approach for genomic prediction in dairy cattle.

Addressing imbalance in graph datasets: Introducing GATE-GNN with graph ensemble weight attention and transfer learning for enhanced node classification

Significant challenges arise when Graph Neural Networks (GNNs) try to deal with uneven data. Specifically in signed and weighted graph structures. This makes classification tasks less effective. Within the GNN context, researchers have found traditional solutions like resampling, reweighting, and synthetic sample generation to be inadequate. GATE-GNN, a novel architecture designed specifically for imbalanced datasets, overcomes these limitations. GATE-GNN integrates an ensemble of network modules that harness the spatial features of graph networks and effectively utilise embedding information from earlier layers. This unique approach not only bolsters generalisation by reducing volatility. It also refines the optimisation algorithm, resulting in more accurate and stable classification outcomes. We rigorously tested the effectiveness of GATE-GNN on four widely recognised datasets: Cora, NELL, Citeseer, and PubMed. We performed a comparative analysis against established methods such as Graph Convolutional Networks (GCN), Graph Sample and Aggregate (GraphSAGE), Propagation Multilayer Perceptron PMLP), Imbalanced Node Sampling GNN (INS-GNN), GNN-Curriculum Learning (GNN-CL) and Graph Attention Networks (GAT). Empirical results demonstrate that GATE-GNN significantly outperforms these existing models, achieving an average improvement in classification accuracy of approximately 5%–10% over the previous best results. Additionally, GATE-GNN presents a marked reduction in training time. This underscores its efficiency and suitability for practical applications in imbalanced graph data scenarios. Implementation of the proposed GATE-GNN can be accessed here https://github.com/afofanah/GATE-GNN.

Simulation-Oriented Analysis and Modeling of Distracted Driving

Distracted driving significantly affects the efficiency and safety of traffic flow. Modeling distracted driving behavior in microscopic traffic flow simulation is essential for understanding its critical impacts on traffic flow. However, due to the influence of various external factors and the considerable uncertainties in behavior characteristics, modeling distracted driving behavior remains a challenge. This study proposed a model which incorporates distraction features into the microscopic traffic flow model to simulate distracted driving behavior. Specifically, the study first examines the characteristics of distracted driving, including the intervals and durations of distraction events, as well as the patterns and environments of distraction. It then introduces distraction parameters into the Intelligent Driver Model (IDM), including reaction time delays and perception deviations in both speed difference and following distance. These parameters are quantified by probabilistic distributions to reflect the uncertainty and individual differences in driving behavior. The model is calibrated and validated using 772 distracted following events from the Shanghai Naturalistic Driving Study (SH-NDS) data. Three patterns of distraction (excessive, moderate, mild) are distinguished and modeled separately. The results show that the model’s accuracy surpasses that of the IDM under various road types and traffic volumes, with an average improvement in model accuracy of about 11.30% on expressways with high traffic volume, 4.54% on expressways with low traffic volume, and 4.46% on surface roads. Meanwhile, the model can effectively simulate the variations in reaction times and perceptual deviations in both speed and following distance for different distraction modes at the individual level, maintaining consistency with reality. Finally, the study simulates distracted driving behavior under different road environments and traffic volumes to explore the impact of distracted driving on traffic flow. The simulation results indicate that an increase in the proportion of distraction reduces the efficiency and safety of traffic flow, which is consistent with real-world observations. Since the model considers human distraction factors, it can generate more dangerous driving scenarios in simulations, which holds significant importance for safety-related research. The findings from this study are expected to be helpful for understanding distracted driving behavior and mitigate its negative influence on the efficiency and safety of traffic flow.

Exploration of MPSO-Two-Stage Classification Optimization Model for Scene Images with Low Quality and Complex Semantics.

Currently, complex scene classification strategies are limited to high-definition image scene sets, and low-quality scene sets are overlooked. Although a few studies have focused on artificially noisy images or specific image sets, none have involved actual low-resolution scene images. Therefore, designing classification models around practicality is of paramount importance. To solve the above problems, this paper proposes a two-stage classification optimization algorithm model based on MPSO, thus achieving high-precision classification of low-quality scene images. Firstly, to verify the rationality of the proposed model, three groups of internationally recognized scene datasets were used to conduct comparative experiments with the proposed model and 21 existing methods. It was found that the proposed model performs better, especially in the 15-scene dataset, with 1.54% higher accuracy than the best existing method ResNet-ELM. Secondly, to prove the necessity of the pre-reconstruction stage of the proposed model, the same classification architecture was used to conduct comparative experiments between the proposed reconstruction method and six existing preprocessing methods on the seven self-built low-quality news scene frames. The results show that the proposed model has a higher improvement rate for outdoor scenes. Finally, to test the application potential of the proposed model in outdoor environments, an adaptive test experiment was conducted on the two self-built scene sets affected by lighting and weather. The results indicate that the proposed model is suitable for weather-affected scene classification, with an average accuracy improvement of 1.42%.

Long-term student performance prediction using learning ability self-adaptive algorithm

Predicting student performance is crucial for both preventing failure and enabling personalized teaching-and-learning strategies. The digitalization of educational institutions has led to the collection of extensive student learning data over the years. Current research primarily focuses on short-term data, e.g. a single year or semester. In contrast, long-term data has the potential to offer a deeper insight into student behavior, thereby increasing the accuracy of predictions. However, the direct application of long-term data in prediction models assumes consistent data distributions over time. In the real world, evolutions in course content and structure can lead to variations in feature spaces (heterogeneity) and distribution shifts across different academic years, compromising the effectiveness of prediction models. To address these challenges, we introduce the Learning Ability Self-Adaptive Algorithm (LASA), which can adapt to the evolving feature spaces and distributions encountered in long-term data. LASA comprises two primary components: Learning Ability Modeling (LAM) and Long-term Distribution Alignment (LTDA). LAM assumes that students’ responses to exercises are samples from distributions that are parameterized by their learning abilities. It then estimates these parameters from the heterogeneous student exercise response data, thereby creating a new homogeneous feature space to counteract the heterogeneity present in long-term data. Subsequently, LTDA employs multiple asymmetric transformations to align distributions of these new features across different years, thus mitigating the impact of distribution shifts on the model’s performance. With these steps, LASA can generate well-aligned features with meaningful semantics. Furthermore, we propose an interpretable prediction framework including three components, i.e. LASA, a base classifier for outcome predictions, and Shapley Additive Explanations (SHAP) for elucidating the impact of specific features on student performance. Our exploration of long-term student data covers an eight-year period (2016-2023) from a face-to-face course at Tsinghua University. Comprehensive experiments demonstrate that leveraging long-term data significantly enhances prediction accuracy compared to short-term data, with LASA achieving up to a 7.9% increase. Moreover, when employing long-term data, LASA outperforms state-of-the-art models, ProbSAP and SFERNN, by an average accuracy improvement of 6.8% and 6.4%, respectively. We also present interpretable insights for pedagogical interventions based on a quantitative analysis of feature impacts on student performance. To the best of our knowledge, this study is the first to investigate student performance prediction in long-term data scenarios, addressing a significant gap in the literature.

Low-resolution few-shot learning via multi-space knowledge distillation

Existing few-shot classification models usually rely on limited known support images to form class centers, and classify query images based on the distance between their embedding and the class centers. However, these models assume that the query image is high-resolution (HR), and thus suffer from significant performance degradation when applied to low-resolution (LR) images. Due to the lack of discriminative information in LR images, there is a noticeable discrepancy between the embeddings of LR query images and the class centers formed by HR support images. To address this issue, we first formulate the problem of Low-Resolution Few-Shot Learning (LRFSL), where the support images are HR while the query images are only available in LR. Then, we propose an end-to-end pipeline that leverages mutual learning between a super-resolution (SR) network and a few-shot classification network. To further reduce the domain discrepancy between the embeddings of the SR images and HR class centers, we introduce a multi-space knowledge distillation strategy that aims to transfer pixel-level, feature-level, and logit-level knowledge of the HR domain to the SR domain. We conduct extensive experiments on classic few-shot datasets: miniImageNet, tieredImageNet, and the fine-grained few-shot dataset CUB. Experimental results show that our method can handle few-shot classification with LR input, and achieve performance that is almost comparable to using HR images as input. Specifically, our method achieves an average accuracy improvement of 26.47% with the Meta-Baseline Model and 7.44% with the Meta DeepBDC Model across all datasets compared to LR Query.

Enhancing target detection accuracy through cross-modal spatial perception and dual-modality fusion

The disparity between human and machine perception of spatial information presents a challenge for machines to accurately sense their surroundings and improve target detection performance. Cross-modal data fusion emerges as a potential solution to enhance the perceptual capabilities of systems. This article introduces a novel spatial perception method that integrates dual-modality feature fusion and coupled attention mechanisms to validate the improvement in detection performance through cross-modal information fusion. The proposed approach incorporates cross-modal feature extraction through a multi-scale feature extraction structure employing a dual-flow architecture. Additionally, a transformer is integrated for feature fusion, while the information perception of the detection system is optimized through the utilization of a linear combination of loss functions. Experimental results demonstrate the superiority of our algorithm over single-modality target detection using visible images, exhibiting an average accuracy improvement of 30.4%. Furthermore, our algorithm outperforms single-modality infrared image detection by 3.0% and comparative multimodal target detection algorithms by 3.5%. These results validate the effectiveness of our proposed algorithm in fusing dual-band features, significantly enhancing target detection accuracy. The adaptability and robustness of our approach are showcased through these results.

An intelligent non-intrusive load monitoring model based on power encoding and convolutional state modules

Abstract Non-intrusive load monitoring (NILM) identifies device power consumption or on/off states solely based on total power data, which is highly valuable for consumers to understand their appliance usage behavior and take necessary measures to reduce energy consumption, especially for the benefit of energy consumers’ living production. However, a challenge faced by NILM is the tendency to focus excessively on power disaggregation while neglecting the disaggregation of on/off states, leading to lower classification accuracy, particularly owning to imbalanced states. This study proposes a model that integrates the power and on/off states to simultaneously disaggregate the power and device on/off states. The model comprises two main modules: a power encoding module for power disaggregation, and a convolutional state module (CSM) for on/off state disaggregation. The power encoding module utilizes BERT-LSTM and long short-term memory networks for initial energy disaggregation. In contrast, the CSM employs convolutional neural networks for device state disaggregation. The output of the power-encoding module is multiplied by the probability of on/off states to obtain the final power. The proposed model is evaluated using the REDD and UK-DALE datasets. Compared to the baseline models, the results show an improvement in the device state classification average accuracy from 0.948 to 0.957, and a decrease in the average error between the real power and disaggregated power from 26.356 W to 25.108 W. Additionally, real-world experiments conducted using the designed platform for collecting and disaggregating power data achieve an average accuracy of 0.997. The proposed model demonstrates competitiveness in the NILM field and underscores its significance in aiding energy-consumption reduction efforts.

A fiber recognition framework based on multi-head attention mechanism

Convolutional neural networks have been extensively studied in textile fiber recognition. However, the down-sampling performed by the convolutional and pooling layers on the extracted features result in a significant loss of fine-grained fiber features. To address this issue, a multi-head attention framework for fiber recognition has been proposed. First, textile surface images are captured using optical magnifiers and smart devices such as smartphones. Second, fiber features and label embeddings are learned separately by multi-head self-attention. Finally, the label embeddings are used to query the presence of fiber types, and pool type-related features in the multi-head cross-attention module to classify fibers. The experimental results demonstrate that the proposed method performs exceptionally well on the textile surface image dataset, with a mean average precision accuracy improvement of 6% compared with the convolutional neural network-based fiber recognition method Cu-Net, and a 5% improvement compared with the fiber recognition method FiberCT combining convolutional neural networks and attention mechanisms. It is worth mentioning that the fiber images captured at 200× magnification in the collected dataset are most favorable for models to recognize fiber. The proposed method achieves a mean average precision fiber recognition accuracy of 80.2% on images at 200× magnification.

Comparing full variation profile analysis with the conventional consensus method in SARS-CoV-2 phylogeny.

This study proposes a novel approach to studying severe acute respiratory syndrome coronavirus 2 virus mutations through sequencing data comparison. Traditional consensus-based methods, which focus on the most common nucleotide at each position, might overlook or obscure the presence of low-frequency variants. Our method, in contrast, retains all sequenced nucleotides at each position, forming a genomic matrix. Utilizing simulated short reads from genomes with specified mutations, we contrasted our genomic matrix approach with the consensus sequence method. Our matrix methodology, across multiple simulated datasets, accurately reflected the known mutations with an average accuracy improvement of 20% over the consensus method. In real-world tests using data from GISAID and NCBI-SRA, our approach demonstrated an increase in reliability by reducing the error margin by approximately 15%. The genomic matrix approach offers a more accurate representation of the viral genomic diversity, thereby providing superior insights into virus evolution and epidemiology.

RFBoost: Understanding and Boosting Deep WiFi Sensing via Physical Data Augmentation

Deep learning shows promising performance in wireless sensing. However, deep wireless sensing (DWS) heavily relies on large datasets. Unfortunately, building comprehensive datasets for DWS is difficult and costly, because wireless data depends on environmental factors and cannot be labeled offline. Despite recent advances in few-shot/cross-domain learning, DWS is still facing data scarcity issues. In this paper, we investigate a distinct perspective of radio data augmentation (RDA) for WiFi sensing and present a data-space solution. Our key insight is that wireless signals inherently exhibit data diversity, contributing more information to be extracted for DWS. We present RFBoost, a simple and effective RDA framework encompassing novel physical data augmentation techniques. We implement RFBoost as a plug-and-play module integrated with existing deep models and evaluate it on multiple datasets. Experimental results demonstrate that RFBoost achieves remarkable average accuracy improvements of 5.4% on existing models without additional data collection or model modifications, and the best-boosted performance outperforms 11 state-of-the-art baseline models without RDA. RFBoost pioneers the study of RDA, an important yet currently underexplored building block for DWS, which we expect to become a standard DWS component of WiFi sensing and beyond. RFBoost is released at https://github.com/aiot-lab/RFBoost.

Interactive attack-defense for generalized person re-identification

Generalized Person Re-Identification (GReID) aims to develop a model capable of robust generalization across unseen target domains, even with training on a limited set of observed domains. Recently, methods based on the Attack-Defense mechanism are emerging as a prevailing technology to this issue, which treats domain transformation as a type of attack and enhances the model’s generalization performance on the target domain by equipping it with a defense module. However, a significant limitation of most existing approaches is their inability to effectively model complex domain transformations, largely due to the separation of attack and defense components. To overcome this limitation, we introduce an innovative Interactive Attack-Defense (IAD) mechanism for GReID. The core of IAD is the interactive learning of two models: an attack model and a defense model. The attack model dynamically generates directional attack information responsive to the current state of the defense model, while the defense model is designed to derive generalizable representations by utilizing a variety of attack samples. The training approach involves a dual process: for the attack model, the aim is to increase the challenge for the defense model in countering the attack; conversely, for the defense model, the focus is on minimizing the effects instigated by the attack model. This interactive framework allows for mutual learning between attack and defense, creating a synergistic learning environment. Our diverse experiments across datasets confirm IAD’s effectiveness, consistently surpassing current state-of-the-art methods, and using MSMT17 as the target domain in different protocols resulted in a notable 13.4% improvement in GReID task average Rank-1 accuracy. Code is available at: https://github.com/lhf12278/IAD.

CamGNN: Cascade Graph Neural Network for Camera Re-Localization

In response to the inaccurate positioning of traditional camera relocation methods in scenes with large-scale or severe viewpoint changes, this study proposes a camera relocation method based on a cascaded graph neural network to achieve accurate scene relocation. Firstly, the NetVLAD retrieval method, which has advantages in image feature representation and similarity calculation, is used to retrieve the most similar images to a given query image. Then, the feature pyramid is employed to extract features at different scales of these images, and the features at the same scale are treated as nodes of the graph neural network to construct a single-layer graph neural network structure. Secondly, a top–down connection is used to cascade the single-layer graph structures, where the information of nodes in the previous graph is fused into a message node to improve the accuracy of camera pose estimation. To better capture the topological relationships and spatial geometric constraints between images, an attention mechanism is introduced in the single-layer graph structure, which helps to effectively propagate information to the next graph during the cascading process, thereby enhancing the robustness of camera relocation. Experimental results on the public dataset 7-Scenes demonstrate that the proposed method can effectively improve the accuracy of camera absolute pose localization, with average translation and rotation errors of 0.19 m and 6.9°, respectively. Compared to other deep learning-based methods, the proposed method achieves more than 10% improvement in both average translation and rotation accuracy, demonstrating highly competitive localization precision.

HybMED: A Hybrid Neural Network Training Processor with Multi-Sparsity Exploitation for Internet of Medical Things.

Cloud-based training and edge-based inference modes for Artificial Intelligence of Medical Things (AIoMT) applications suffer from accuracy degradation due to physiological signal variations among patients. On-chip learning can overcome this issue by online adaptation of neural network parameters for user-specific tasks. However, existing on-chip learning processors have limitations in terms of versatility, resource utilization, and energy efficiency. We propose HybMED, which is a novel neural signal processor that supports on-chip hybrid neural network training using a composite direct feedback alignment-based paradigm. HybMED is suitable for general-purpose health monitoring AIoMT devices. It improves resource utilization and area efficiency by the reconfigurable homogeneous core with heterogeneous data flow and enhances energy efficiency by exploiting sparsity at different granularities. The chip was fabricated by TSMC 40nm process and tested in multiple physiological signal processing tasks, demonstrating an average improvement in accuracy of 41.16% following online few-shot learning. The chip demonstrates an area efficiency of 1.17 GOPS/mm2 and an energy efficiency of 1.58 TOPS/W. Compared to the previous state-of-the-art physiological signal processors with on-chip learning, the chip achieves a 65× improvement in area efficiency and 1.48× improvement in energy efficiency, respectively.

Feature decoupling and regeneration towards wifi-based human activity recognition

Gesture recognition using WiFi is vital for human–computer interaction, smart homes, smart spaces, and elderly care. WiFi signals are non-stationary, sensitive to the environment, and traditional pattern recognition-based HAR methods incur exorbitant training and deployment costs due to their reliance on the quality and quantity of training data. While there has been extensive research on enhancing HAR techniques, the efficiency of the system is still limited by the scarcity of training data and identity-action feature coupling. This research focuses on attaining gesture recognition across multiple users with a limited number of samples. Our findings show that the conventional data augmentation methods are incapable of facilitating the system in attaining sample diversity. As a result, we propose a training-free augmentation strategy as a means to provide adequate training data. Unlike the conventional data enhancement approach, this scheme aims to develop data processing methods that differentiate between various samples in practical applications. Consequently, it enhances data pertinent to specific HAR assignments effectively. To effectively extract action features in WiFi samples, an unsupervised cross-user domain sample generation (CUDSG) model is proposed. This model generates virtual gesture samples for new user domains by decoupling and recombining gesture and identity features. This extends the sensing boundary of the system to new user domains without requiring a significant number of users to participate. Model performance was evaluated using various classifiers, such as SVM, KNN and CNN. The results demonstrate a significant improvement in average classification accuracy from 57.3% to 98.4%. This indicates that CUDSG is a highly effective tool for enhancing the performance of existing gesture recognition techniques.