2787 papers
This paper proposes a sequence-local expert roster strategy that fixes a small set of experts per sequence to drastically reduce memory footprint and expert transfers during MoE model serving, achieving near-full-residence throughput with significantly fewer resident weights by applying this constraint consistently during both training and inference.
This paper introduces a novel homogeneity metric to evaluate the effectiveness and efficiency of unsupervised multivariate time series segmentation algorithms, demonstrating their successful application on both synthetic data and real-world automotive embedded system data.
This paper proposes a unified supervised multi-view subspace learning method that integrates class-wise geometric structures and adaptive view weighting to efficiently construct a discriminative joint latent space, outperforming state-of-the-art approaches in both accuracy and computational speed.
This paper proposes a hybrid deep-quantum framework that integrates a 1D-CNN for feature extraction with a quantum support vector classifier to achieve superior stress detection accuracy (81.00%) on EEG signals compared to classical baselines, while demonstrating the model's dependence on feature scaling and dimensionality within the NISQ era.
This paper proposes WTFFN-YOLO, an improved lightweight object detection framework for wind turbine blade surface defects that integrates the C3K2-SACA attention mechanism, a frequency-aware feature fusion module, and a WTFFN neck network to achieve a 6.9% increase in mAP and an 8% reduction in model parameters.
This study demonstrates that a hybrid CNN–BiLSTM architecture outperforms traditional acoustic feature analysis and simpler deep learning models in quantifying linguistic rhythm by achieving over 95% accuracy in distinguishing Sanskrit and Hindi chanting from normal speech, while revealing greater rhythmic regularity in Sanskrit due to its syllable-timed structure.
This paper reveals that Physics-Informed Neural Networks (PINNs) applied to SIR models suffer from "latent debt" during early exponential growth phases, where a flat, rank-deficient loss landscape allows optimization to minimize residuals by distorting unobserved latent states and propagating errors, ultimately yielding numerically accurate solutions with fundamentally incorrect parameter estimates.
This study demonstrates the significant performance advantages of NVIDIA RTX 3050 GPUs over Intel CPUs in deep learning training, revealing a 40.6-fold speedup for large-scale ResNet50 models on the CIFAR-10 dataset.
Xender+ is a novel framework that enhances time-series data transmission efficiency by replacing frequent online model retraining with a cloud-side library of pre-trained, pattern-clustered models for direct reuse, thereby achieving lower end-to-end latency and high reconstruction accuracy compared to state-of-the-art baselines.
This paper introduces a resource-aware benchmarking framework utilizing adaptive temporal sampling to evaluate efficiency–accuracy trade-offs in video action recognition, demonstrating that the R(2+1)D-18 architecture significantly reduces energy consumption while maintaining stable performance across varying computational budgets.