Detecting rotated faces has always been a challenging task. Fixed convolutional kernels struggle to effectively match features after rotation, while the sampling point offsets of deformable convolutions are limited by...
详细信息
The Processing-In-Memory (PIM) architecture becomes a promising candidate for deep learning accelerators by integrating computation and memory. Most PIM-based studies improve the performance and energy efficiency by u...
详细信息
The Processing-In-Memory (PIM) architecture becomes a promising candidate for deep learning accelerators by integrating computation and memory. Most PIM-based studies improve the performance and energy efficiency by using the Weight Stationary (WS) data flow due to its high parallelism. However, the WS data flow has some fundamental limitations. First, the WS data flow has huge activation movements between on-chip memory and off-chip memory due to the limited memory space of the ReRAM array. Second, the WS data flow needs to read the input activation repeatedly according to the convolution window. These data movements decrease the energy efficiency and performance of the PIM architecture. To address these issues, the IS data flow stores activations instead of weights to reduce data movements. But the IS data flow faces some challenges. First, the data dependency between adjacent layers limits the performance. Second, there are huge across-array computations due to the special mapping method. Third, the previous IS data flow cannot realize the high parallelism. Fourth, the IS data flow depends on the three-dimensional (3D) ReRAM structure. To address these issues, we propose a novel data flow for PIM architectures. We optimize the IS data flow to decrease the activation movement and propose a parallel computing method to realize high parallelism and reduce the across-array computations. We identify and analyze the fundamental limitations and impact of different inter-layer data flows, including the WS-WS, IS-IS, WS-IS, and IS-WS. We also propose a method to build a hybrid data flow by combining these inter-layer data flows to trade-off performance and energy consumption. Our experimental results and analysis demonstrate the potential of our design. The performance and energy efficiency of our design reaches 0.13 TFLOPS∼1.77 TFLOPS and 61 TOPS/J∼85 TOPS/J, respectively. Compared to the state-of-the-art design, the NEBULA, our design can improve performance by 1.4×, 2.
State-of-the-art recommender systems are increasingly focused on optimizing implementation efficiency, such as enabling on-device recommendations under memory constraints. Current methods commonly use lightweight embe...
详细信息
State-of-the-art recommender systems are increasingly focused on optimizing implementation efficiency, such as enabling on-device recommendations under memory constraints. Current methods commonly use lightweight embeddings for users and items or employ compact embeddings to enhance reusability and reduce memory usage. However, these approaches consider only the coarse-grained aspects of embeddings, overlooking subtle semantic nuances. This limitation results in an adversarial degradation of meta-embedding performance, impeding the system's ability to capture intricate relationships between users and items, leading to suboptimal recommendations. To address this, we propose a novel approach to efficiently learn meta-embeddings with varying grained and apply fine-grained meta-embeddings to strengthen the representation of their coarse-grained counterparts. Specifically, we introduce a recommender system based on a graph neural network, where each user and item is represented as a node. These nodes are directly connected to coarse-grained virtual nodes and indirectly linked to fine-grained virtual nodes, facilitating learning of multi-grained semantics. Fine-grained semantics are captured through sparse meta-embeddings, which dynamically balance embedding uniqueness and memory constraints. To ensure their sparseness, we rely on initialization methods such as sparse principal component analysis combined with a soft thresholding activation function. Moreover, we propose a weight-bridging update strategy that aligns coarse-grained meta-embedding with several fine-grained meta-embeddings based on the underlying semantic properties of users and items. Comprehensive experiments demonstrate that our method outperforms existing baselines. The code of our proposal is available at https://***/htyjers/C2F-MetaEmbed.
WiFi-based indoor positioning has emerged as a crucial technology for enabling smart consumer electronic applications, particularly in large-scale buildings. The construction of WiFi fingerprint databases using receiv...
详细信息
Long-term multivariate time series forecasting is an important task in engineering applications. It helps grasp the future development trend of data in real-time, which is of great significance for a wide variety of f...
详细信息
Long-term multivariate time series forecasting is an important task in engineering applications. It helps grasp the future development trend of data in real-time, which is of great significance for a wide variety of fields. Due to the non-linear and unstable characteristics of multivariate time series, the existing methods encounter difficulties in analyzing complex high-dimensional data and capturing latent relationships between multivariates in time series, thus affecting the performance of long-term prediction. In this paper, we propose a novel time series forecasting model based on multilayer perceptron that combines spatio-temporal decomposition and doubly residual stacking, namely Spatio-Temporal Decomposition Neural Network (STDNet). We decompose the originally complex and unstable time series into two parts, temporal term and spatial term. We design temporal module based on auto-correlation mechanism to discover temporal dependencies at the sub-series level, and spatial module based on convolutional neural network and self-attention mechanism to integrate multivariate information from two dimensions, global and local, respectively. Then we integrate the results obtained from the different modules to get the final forecast. Extensive experiments on four real-world datasets show that STDNet significantly outperforms other state-of-the-art methods, which provides an effective solution for long-term time series forecasting.
Metapaths with specific complex semantics are critical to learning diverse semantic and structural information of heterogeneous networks(HNs)for most of the existing representation learning ***,any metapaths consistin...
详细信息
Metapaths with specific complex semantics are critical to learning diverse semantic and structural information of heterogeneous networks(HNs)for most of the existing representation learning ***,any metapaths consisting of multiple,simple metarelations must be driven by domain *** sensitive,expensive,and limited metapaths severely reduce the flexibility and scalability of the existing models.A metapath-free,scalable representation learning model,called Metarelation2vec,is proposed for HNs with biased joint learning of all metarelations in a bid to address this ***,a metarelation-aware,biased walk strategy is first designed to obtain better training samples by using autogenerating cooperation probabilities for all metarelations rather than using expert-given ***,grouped nodes by the type,a common and shallow skip-gram model is used to separately learn structural proximity for each node ***,grouped links by the type,a novel and shallow model is used to separately learn the semantic proximity for each link ***,supervised by the cooperation probabilities of all meta-words,the biased training samples are thrown into the shallow models to jointly learn the structural and semantic information in the HNs,ensuring the accuracy and scalability of the *** experimental results on three tasks and four open datasets demonstrate the advantages of our proposed model.
Brain tumor classification is crucial for personalized treatment *** deep learning-based Artificial Intelligence(AI)models can automatically analyze tumor images,fine details of small tumor regions may be overlooked d...
详细信息
Brain tumor classification is crucial for personalized treatment *** deep learning-based Artificial Intelligence(AI)models can automatically analyze tumor images,fine details of small tumor regions may be overlooked during global feature ***,we propose a brain tumor Magnetic Resonance Imaging(MRI)classification model based on a global-local parallel dual-branch *** global branch employs ResNet50 with a Multi-Head Self-Attention(MHSA)to capture global contextual information from whole brain images,while the local branch utilizes VGG16 to extract fine-grained features from segmented brain tumor *** features from both branches are processed through designed attention-enhanced feature fusion module to filter and integrate important ***,to address sample imbalance in the dataset,we introduce a category attention block to improve the recognition of minority *** results indicate that our method achieved a classification accuracy of 98.04%and a micro-average Area Under the Curve(AUC)of 0.989 in the classification of three types of brain tumors,surpassing several existing pre-trained Convolutional Neural Network(CNN)***,feature interpretability analysis validated the effectiveness of the proposed *** suggests that the method holds significant potential for brain tumor image classification.
Stock price prediction is a typical complex time series prediction problem characterized by dynamics,nonlinearity,and *** paper introduces a generative adversarial network model that incorporates an attention mechanis...
详细信息
Stock price prediction is a typical complex time series prediction problem characterized by dynamics,nonlinearity,and *** paper introduces a generative adversarial network model that incorporates an attention mechanism(GAN-LSTM-Attention)to improve the accuracy of stock price ***,the generator of this model combines the Long and Short-Term Memory Network(LSTM),the Attention Mechanism and,the Fully-Connected Layer,focusing on generating the predicted stock *** discriminator combines the Convolutional Neural Network(CNN)and the Fully-Connected Layer to discriminate between real stock prices and generated stock ***,to evaluate the practical application ability and generalization ability of the GAN-LSTM-Attention model,four representative stocks in the United States of America(USA)stock market,namely,Standard&Poor’s 500 Index stock,Apple Incorporatedstock,AdvancedMicroDevices Incorporatedstock,and Google Incorporated stock were selected for prediction experiments,and the prediction performance was comprehensively evaluated by using the three evaluation metrics,namely,mean absolute error(MAE),root mean square error(RMSE),and coefficient of determination(R2).Finally,the specific effects of the attention mechanism,convolutional layer,and fully-connected layer on the prediction performance of the model are systematically analyzed through ablation *** results of experiment show that the GAN-LSTM-Attention model exhibits excellent performance and robustness in stock price prediction.
In low-light image enhancement,prevailing Retinex-based methods often struggle with precise illumina-tion estimation and brightness *** can result in issues such as halo artifacts,blurred edges,and diminished details ...
详细信息
In low-light image enhancement,prevailing Retinex-based methods often struggle with precise illumina-tion estimation and brightness *** can result in issues such as halo artifacts,blurred edges,and diminished details in bright regions,particularly under non-uniform illumination *** propose an innovative approach that refines low-light images by leveraging an in-depth awareness of local content within the *** introducing multi-scale effective guided filtering,our method surpasses the limitations of traditional isotropic filters,such as Gaussian filters,in handling non-uniform *** dynamically adjusts regularization parameters in response to local image characteristics and significantly integrates edge perception across different *** balanced approach achieves a harmonious blend of smoothing and detail preservation,enabling more accurate illumination ***,we have designed an adaptive gamma correction function that dynamically adjusts the brightness value based on local pixel intensity,further balancing enhancement effects across different brightness levels in the *** results demonstrate the effectiveness of our proposed method for non-uniform illumination images across various *** exhibits superior quality and objective evaluation scores compared to existing *** method effectively addresses potential issues that existing methods encounter when processing non-uniform illumination images,producing enhanced images with precise details and natural,vivid colors.
Graph similarity learning aims to calculate the similarity between pairs of *** unsupervised graph similarity learning methods based on contrastive learning encounter challenges related to random graph augmentation st...
详细信息
Graph similarity learning aims to calculate the similarity between pairs of *** unsupervised graph similarity learning methods based on contrastive learning encounter challenges related to random graph augmentation strategies,which can harm the semantic and structural information of graphs and overlook the rich structural information present in *** address these issues,we propose a graph similarity learning model based on learnable augmentation and multi-level contrastive ***,to tackle the problem of random augmentation disrupting the semantics and structure of the graph,we design a learnable augmentation method to selectively choose nodes and edges within the *** enhance contrastive levels,we employ a biased random walk method to generate corresponding subgraphs,enriching the contrastive ***,to solve the issue of previous work not considering multi-level contrastive learning,we utilize graph convolutional networks to learn node representations of augmented views and the original graph and calculate the interaction information between the attribute-augmented and structure-augmented views and the original *** goal is to maximize node consistency between different views and learn node matching between different graphs,resulting in node-level representations for each *** representations are then obtained through pooling operations,and we conduct contrastive learning utilizing both node and subgraph ***,the graph similarity score is computed according to different downstream *** conducted three sets of experiments across eight datasets,and the results demonstrate that the proposed model effectively mitigates the issues of random augmentation damaging the original graph’s semantics and structure,as well as the insufficiency of contrastive ***,the model achieves the best overall performance.
暂无评论