Most state-of-the-art convolutional neural networks (CNNs) are characterized by excessive parameterization, leading to a high computational burden. Tensor decomposition has emerged as a model reduction technique for compressing deep neural networks. Previous approaches have predominantly relied on either Tucker decomposition or Canonical Polyadic (CP) decomposition for CNNs. However, CP decomposition exhibits exceptional compression capabilities in comparison to Tucker decomposition, which results in a more pronounced accuracy loss. This article introduces an efficient model compression method, termed TEC-CNN, designed to achieve significant compression while preserving accuracy levels comparable to those of the original models. In TEC-CNN, convolutional layers are identified to obtain convolutional kernels by analyzing given models under the principles of low-rank tensor decomposition, and then, calculating the ranks of convolutional kernels. Furthermore, an efficient decomposition schema for the convolutional kernel is proposed with approximate kernel tensor for reducing parameters. Additionally, a novel format of a convolutional sequence is presented and constructed with a reduced number of parameters to replace the original convolutional layers. Finally, the effectiveness of TEC-CNN is assessed across a range of computer vision tasks. For instance, in CIFAR-100 classification, ResNet18 is compressed to 4.1 MB, while Unext, when applied to image segmentation using the International Skin Imaging Collaboration (ISIC) dataset, is reduced to 3.419 MB. When employed for fire object detection with Yolov7, TEC-CNN achieves a model size reduction of 71.6 MB. Comprehensive experimental results underscore that our approach achieves significant model compression while preserving model performance.

Background: Although much progress has been made in artificial intelligence (AI), several challenges remain substantial obstacles to the development and translation of AI systems into clinical practice. Even large language models, which show excellent performance on various tasks, have progressed slowly in clinical practice tasks. Providing precise and explainable treatment plans with personalized details remains a big challenge for AI systems due to both the highly specialized medical knowledge required and patients’ complicated conditions. Objective: This study aimed to develop an explainable and efficient decision support system named H-SYSTEM to assist neurosurgeons in diagnosing and treating patients with hypertensive intracerebral hemorrhage. The system was designed to address the limitations of existing AI systems by integrating a medical domain knowledge graph to enhance decision-making accuracy and explainability. Methods: The H-SYSTEM consists of 3 main modules: the key named entity recognition (NER) module, the semantic analysis and representation module, and the reasoning module. Furthermore, we constructed a medical domain knowledge graph for hypertensive intracerebral hemorrhage, named HKG, which served as an external knowledge brain of the H-SYSTEM to enhance its text recognition and automated decision-making capability. The HKG was exploited to guide the training of the semantic analysis and representation module and reasoning module, which makes the output of the H-SYSTEM more explainable. To assess the performance of the H-SYSTEM, we compared it with doctors and different large language models. Results: The outputs based on HKG showed reliable performance as compared with neurosurgical doctors, with an overall accuracy of 94.87%. The bidirectional encoder representations from transformers, inflated dilated convolutional neural network, bidirectional long short-term memory, and conditional random fields (BERT-IDCNN-BiLSTM-CRF) model was used as the key NER module of the H-SYSTEM due to its fast convergence and efficient extraction of key named entities, achieved the highest performance among 7 key NER models (precision=92.03, recall=90.22, and F₁-score=91.11), significantly outperforming the others. The H-SYSTEM achieved an overall accuracy of 91.74% in treatment plans, showing significant consistency with the gold standard (P<.05), with diagnostic measures achieving 88.18% accuracy, 97.03% area under the curve (AUC), and a κ of 0.874; surgical therapy achieving 98.53% accuracy, 98.53% AUC, and a κ of 0.971; and rescue therapies achieving 89.50% accuracy, 94.67% AUC, and a κ of 0.923 (all P<.05). Furthermore, the H-SYSTEM showed high reliability and efficiency when compared to doctors and ChatGPT, achieving statistically higher accuracy (95.26% vs 91.48%, P<.05). Additionally, the H-SYSTEM achieved a total accuracy of 92.22% (ranging from 91.14% to 95.35%) in treatment plans for 605 additional patients from 6 different medical centers. Conclusions: The H-SYSTEM showed significantly high efficiency and generalization capacity in processing electronic medical records, and it provided explainable and elaborate treatment plans. Therefore, it has the potential to provide neurosurgeons with rapid and reliable decision support, especially in emergency conditions. The knowledge graph–enhanced deep-learning model exhibited excellent performance in the clinical practice tasks.

Efficient load management and system stability in active distribution networks (ADNs) depend heavily on accurate load forecasting and dynamic decision-making. This paper proposes a novel end-to-end joint learning model for dynamic prediction-decision joint-making in active distribution networks. The model integrates a prediction module, leveraging long short-term memory networks to forecast photovoltaic (PV) generation, active power load, and reactive power load, with a control decision module based on Twin Delayed Deep Deterministic Policy Gradient (TD3). The forecasting outputs are fed into the decision-making module, enabling dynamic adjustments to network parameters. A feedback mechanism is introduced that incorporates decision losses into the prediction module to enhance forecasting accuracy, thereby creating a tightly coupled, end-to-end joint optimization framework. In addition, a mixed precision quantization technology is adopted to reduce computational overhead and facilitate efficient deployment. Experimental results on distribution networks demonstrate that the proposed model outperforms other methods in terms of load forecasting accuracy and dynamic control effectiveness, resulting in substantial improvements in the operational efficiency and stability of ADNs.

Convolutional Neural Networks (CNNs) have achieved significant success in feature representation. However, their feature representation capability can be affected by high crossover points between classes within datasets, resulting in non-Gaussian and heterogeneous distributions. To address this, we propose a novel feature representation method called CNN-MGDNF, which captures features through convolution and pooling operations. A maximum Gaussianity normalizing flow (MGDNF) model is designed to transform the original feature space into a Gaussian space, balancing intervals between classes. Additionally, an entropy-based imbalance degree (EID) is explored to mitigate multi-class imbalance issues. Extensive experiments on the yak, vehicle identification, and bird species datasets demonstrate that CNN-MGDNF enhances feature representation, achieving remarkable accuracies of 99.59%, 97.93%, and 90.39%, respectively. Our method significantly outperforms traditional CNN models and other methods, underscoring its effectiveness in visual recognition tasks.

The huge computational complexity and memory requirement of deep neural networks (DNNs) impede their deployability on resource-constrained edge-computing devices. To tackle this challenge, numerous model compression solutions have emerged, where Power-of-Two (PoT) quantization scheme demonstrates its potential in reducing computational complexity with complex multiplications replaced by bit-shifting arithmetic operations. However, most of existing methods resort to using full-precision hyper-parameters to avoid significant accuracy degradation, and so they achieve sub-optimal performance on the reduction of computation and memory cost. Therefore, we propose a novel efficient inference approach for resource-limited embedded devices. In our approach, integer-only PoT quantization scheme is designed to quantize both weights and activations into only PoT representations with fewer bit-width to achieve bit-wise inference. Additionally, a distribution-loss regularizer is developed to minimize the retraining disturbation introduced by quantizer. Furthermore, output weights are efficiently encoded in binary format by a two-stage compression pipeline. Finally, comprehensive experiments are conducted with different models and datasets. Experimental results indicate that our approach outperforms state-of-the-art approaches in terms of accuracy and compression rate, with more than 2 × the compression ratio previous methods achieve and 1.2 × accuracy improvement, under 3-bit precision with ResNet56.