Min Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Min Liu has a total of 13 co-authors. Here are the top co-authors by number of shared publications:
- Zhe Li
Department of Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China
8 shared publications - Gehao Sheng
Department of Electrical Engineering, Shanghai Jiao Tong University, Shanghai 200240, China
5 shared publications - Yaocheng Li
China Electric Power Research Institute, Beijing, China
4 shared publications - Ran Tao
Department of Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China
3 shared publications - Xiuchen Jiang
School of Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China
2 shared publications - Changchun Zhou
Duke University, Durham, NC, USA
1 shared publication
Latest Publications
- 11 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dielectrics and Electrical Insulation
- Vol. PP (10709878) , 1
A ML-Driven Study on Feature Descriptors and Mechanisms for Adsorption of Air Discharge Decomposition Components
Insulation materials in air switchgear may degrade due to partial discharge (PD). Characteristic gases generated by PD (such as CO, NO, and NO₂) can be detected by gas sensors. However, existing gas-sensitive material screening relies on trial-and-error methods or costly density functional theory (DFT) calculations. Machine learning (ML)-driven DFT calculations have emerged as a new paradigm in gas-sensitive material research. This approach significantly accelerates the rational design of high-performance gas-sensitive materials by efficiently predicting adsorption behavior and revealing underlying microscopic mechanisms. This study screened 14 key features from 200 sets of adsorption data across 50 literature sources (2018–2024), then compared and optimized six ML models (XGBoost, GBR, etc.) to evaluate the predictive performance of D and Ead. Subsequently, SHAP and PDPs analyses were conducted to quantify feature contributions and identify core factors. Research findings indicate that XGBoost achieved prediction R² values of 91.55% for adsorption distance and 89.79% for adsorption energy, significantly outperforming other models (GBR ranked second with R² values of 89.79% and 83.56%, respectively). In terms of generalization ability, XGBoost outperforms GBR in extreme value usage and feature balancing utilization, which is overly dependent on substrate material features. Charge transfer (Qₜ) and dopant atom density (ρd) are the core factors affecting the adsorption behavior, and the SHAP analysis shows that Qₜ has the highest contribution to the model output. Substrate and dopant characteristics contribute equally in the XGBoost model (∼45% each), while the GBR model is overly dependent on substrate characteristics (>60%).
- 11 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dielectrics and Electrical Insulation
- Vol. PP (10709878) , 1
A ML-Driven Study on Feature Descriptors and Mechanisms for Adsorption of Air Discharge Decomposition Components
Insulation materials in air switchgear may degrade due to partial discharge (PD). Characteristic gases generated by PD (such as CO, NO, and NO₂) can be detected by gas sensors. However, existing gas-sensitive material screening relies on trial-and-error methods or costly density functional theory (DFT) calculations. Machine learning (ML)-driven DFT calculations have emerged as a new paradigm in gas-sensitive material research. This approach significantly accelerates the rational design of high-performance gas-sensitive materials by efficiently predicting adsorption behavior and revealing underlying microscopic mechanisms. This study screened 14 key features from 200 sets of adsorption data across 50 literature sources (2018–2024), then compared and optimized six ML models (XGBoost, GBR, etc.) to evaluate the predictive performance of D and Ead. Subsequently, SHAP and PDPs analyses were conducted to quantify feature contributions and identify core factors. Research findings indicate that XGBoost achieved prediction R² values of 91.55% for adsorption distance and 89.79% for adsorption energy, significantly outperforming other models (GBR ranked second with R² values of 89.79% and 83.56%, respectively). In terms of generalization ability, XGBoost outperforms GBR in extreme value usage and feature balancing utilization, which is overly dependent on substrate material features. Charge transfer (Qₜ) and dopant atom density (ρd) are the core factors affecting the adsorption behavior, and the SHAP analysis shows that Qₜ has the highest contribution to the model output. Substrate and dopant characteristics contribute equally in the XGBoost model (∼45% each), while the GBR model is overly dependent on substrate characteristics (>60%).
- 24 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 1-13
Self-Supervised Small-Object Detection YOLO for Power Transmission Line Inspection
Power line inspection utilizing UAV-mounted vision-based measurement systems enables real-time condition monitoring and defect detection of transmission lines, ensuring safe power system operation. However, current power line defect detection algorithms underutilize vast amounts of unlabeled inspection images and struggle with detecting small defects. To this end, a novel Self-supervised Small object detection YOLO (SS-YOLO) is proposed. SS-YOLO consists of a self-supervised pretraining stage using an improved masked image modeling (MIM) and a finetuning stage on an improved YOLOv11 model. First, for the improved MIM, a sparse backbone is introduced to mitigate the masked pattern vanishing problem in the traditional YOLO backbone, the YOLO neck is incorporated to enable transfer learning, and an easy-to-hard masking strategy is suggested to help the model reconstruct transmission line foregrounds, resulting in more effective pretraining. Second, for the improved YOLOv11 model, re-parameterized complementary module (RepCM) is presented to complement positional and semantic features of small objects, re-parameterized perception module (RepPM) is developed to achieve multi-kernel perception of small object features, and normalized Wasserstein distance (NWD) metric is utilized for label assignment and box loss, which improves both accuracy and speed for small defect detection. Experimental results demonstrate that SS-YOLO achieves a 10.9% improvement in power line small defect detection accuracy compared to YOLOv11 and realizes efficient detection on edge computing devices.
- 24 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 1-13
Self-Supervised Small-Object Detection YOLO for Power Transmission Line Inspection
Power line inspection utilizing UAV-mounted vision-based measurement systems enables real-time condition monitoring and defect detection of transmission lines, ensuring safe power system operation. However, current power line defect detection algorithms underutilize vast amounts of unlabeled inspection images and struggle with detecting small defects. To this end, a novel Self-supervised Small object detection YOLO (SS-YOLO) is proposed. SS-YOLO consists of a self-supervised pretraining stage using an improved masked image modeling (MIM) and a finetuning stage on an improved YOLOv11 model. First, for the improved MIM, a sparse backbone is introduced to mitigate the masked pattern vanishing problem in the traditional YOLO backbone, the YOLO neck is incorporated to enable transfer learning, and an easy-to-hard masking strategy is suggested to help the model reconstruct transmission line foregrounds, resulting in more effective pretraining. Second, for the improved YOLOv11 model, re-parameterized complementary module (RepCM) is presented to complement positional and semantic features of small objects, re-parameterized perception module (RepPM) is developed to achieve multi-kernel perception of small object features, and normalized Wasserstein distance (NWD) metric is utilized for label assignment and box loss, which improves both accuracy and speed for small defect detection. Experimental results demonstrate that SS-YOLO achieves a 10.9% improvement in power line small defect detection accuracy compared to YOLOv11 and realizes efficient detection on edge computing devices.
- 4 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. PP (00189456) , 1
KCDet: Keypoint Clustering Detector for Small Bolt Defect Detection in Power Line Inspection
The UAV-based power line intelligent inspection using vision-based instrumentation and measurement plays a crucial role in ensuring the safety of transmission lines. However, small object defects, such as bolt defects, are difficult to detect due to their minimal size in captured images, resulting in low detection rates. To address this issue, this paper proposes a novel keypoint clustering detector (KCDet), consisting of a keypoint detection network, an efficient clustering method, and an object detector, to accurately identify bolt defects. Firstly, a keypoint detection network based on the high-resolution heatmap is proposed, treating bolts as points for localization, which significantly improves the accuracy of bolt localization. Secondly, an efficient clustering method is introduced to cluster the localized bolts, and the cluster regions are cropped from the original image to obtain high-resolution sub-images. Finally, an improved YOLOv11 is presented to precisely identify bolt defects from these sub-images. Experimental results demonstrate that the proposed KCDet increases accuracy by over 20% compared to the single-model-based YOLOv11s algorithm while maintaining high inference speed, significantly enhancing the accuracy of small-sized bolt defect detection in power transmission line inspections.
- 4 September 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. PP (00189456) , 1
KCDet: Keypoint Clustering Detector for Small Bolt Defect Detection in Power Line Inspection
The UAV-based power line intelligent inspection using vision-based instrumentation and measurement plays a crucial role in ensuring the safety of transmission lines. However, small object defects, such as bolt defects, are difficult to detect due to their minimal size in captured images, resulting in low detection rates. To address this issue, this paper proposes a novel keypoint clustering detector (KCDet), consisting of a keypoint detection network, an efficient clustering method, and an object detector, to accurately identify bolt defects. Firstly, a keypoint detection network based on the high-resolution heatmap is proposed, treating bolts as points for localization, which significantly improves the accuracy of bolt localization. Secondly, an efficient clustering method is introduced to cluster the localized bolts, and the cluster regions are cropped from the original image to obtain high-resolution sub-images. Finally, an improved YOLOv11 is presented to precisely identify bolt defects from these sub-images. Experimental results demonstrate that the proposed KCDet increases accuracy by over 20% compared to the single-model-based YOLOv11s algorithm while maintaining high inference speed, significantly enhancing the accuracy of small-sized bolt defect detection in power transmission line inspections.
- 24 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dielectrics and Electrical Insulation
- Vol. 32 (6) , 3688-3697
Defect Detection of Power Line Insulator Image Based on Self-Supervised Pretraining and YOLOv11
Insulators are crucial components in power transmission lines, serving essential insulation functions. UAV-based intelligent inspection significantly enhances the efficiency of insulator defect detection via image recognition. However, the scarcity of labeled insulator defect data hinders training, the small size of defect targets reduces recognition accuracy, and the computational complexity of recognition models poses challenges for drone deployment. To address these issues, firstly, a self-supervised pretraining method termed Insu-MIM is proposed, which uses a large amount of unlabeled insulator data to pretrain the encoder and decoder, acquiring prior knowledge transferable to downstream detection tasks, thereby improving small-sample generalization. Furthermore, an improved YOLOv11 is presented by adding a small object detection layer and incorporating the NWD loss function, enhancing the model’s ability to detect small insulator defects. Lastly, KD-GCP model compression method is suggested, which employs group channel pruning to remove redundant channels and incorporates knowledge distillation to improve the accuracy of the pruned model. Experimental results demonstrate that our proposed approach increases the insulator defect recognition accuracy by 9.5% compared to the original YOLOv11 method, while reducing computational cost by 43%, significantly enhancing the accuracy and efficiency of insulator inspections in power transmission lines.
- 24 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Dielectrics and Electrical Insulation
- Vol. 32 (6) , 3688-3697
Defect Detection of Power Line Insulator Image Based on Self-Supervised Pretraining and YOLOv11
Insulators are crucial components in power transmission lines, serving essential insulation functions. UAV-based intelligent inspection significantly enhances the efficiency of insulator defect detection via image recognition. However, the scarcity of labeled insulator defect data hinders training, the small size of defect targets reduces recognition accuracy, and the computational complexity of recognition models poses challenges for drone deployment. To address these issues, firstly, a self-supervised pretraining method termed Insu-MIM is proposed, which uses a large amount of unlabeled insulator data to pretrain the encoder and decoder, acquiring prior knowledge transferable to downstream detection tasks, thereby improving small-sample generalization. Furthermore, an improved YOLOv11 is presented by adding a small object detection layer and incorporating the NWD loss function, enhancing the model’s ability to detect small insulator defects. Lastly, KD-GCP model compression method is suggested, which employs group channel pruning to remove redundant channels and incorporates knowledge distillation to improve the accuracy of the pruned model. Experimental results demonstrate that our proposed approach increases the insulator defect recognition accuracy by 9.5% compared to the original YOLOv11 method, while reducing computational cost by 43%, significantly enhancing the accuracy and efficiency of insulator inspections in power transmission lines.
- 21 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-11
A Deep Learning Algorithm for Power Transformer Voiceprint Recognition in Strong-Noise and Small-Sample Scenarios
Transformer voiceprint recognition, which involves the analysis of transformer sounds to detect potential faults, has become a research hot spot in recent years. However, sound signals are susceptible to noise interference, and obtaining enough labeled transformer sound samples is challenging, thereby severely impacting the recognition accuracy. To this end, a deep learning algorithm for power transformer voiceprint recognition in strong-noise and small-sample scenarios is offered in this article. First, based on the frequency distribution characteristics of transformer body sounds, an improved spectrum method is proposed, selectively extracting transformer body frequency while suppressing environmental noise frequency. Second, ResVGG transformer voiceprint classification network is presented, which decouples training and inference to simultaneously enhance training accuracy and inference speed. Furthermore, a denoising network is introduced as a precursor to the classification network to further eliminate noise from the spectrum. Finally, self-supervised contrastive learning is employed, utilizing a large number of unlabeled sound samples for pretraining to enhance small-sample generalization and noise robustness. Experimental results demonstrate that the proposed method achieves a 10.6% improvement in accuracy compared to the baseline method, while also exhibiting satisfactory performance in inference time and memory usage.
- 21 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-11
A Deep Learning Algorithm for Power Transformer Voiceprint Recognition in Strong-Noise and Small-Sample Scenarios
Transformer voiceprint recognition, which involves the analysis of transformer sounds to detect potential faults, has become a research hot spot in recent years. However, sound signals are susceptible to noise interference, and obtaining enough labeled transformer sound samples is challenging, thereby severely impacting the recognition accuracy. To this end, a deep learning algorithm for power transformer voiceprint recognition in strong-noise and small-sample scenarios is offered in this article. First, based on the frequency distribution characteristics of transformer body sounds, an improved spectrum method is proposed, selectively extracting transformer body frequency while suppressing environmental noise frequency. Second, ResVGG transformer voiceprint classification network is presented, which decouples training and inference to simultaneously enhance training accuracy and inference speed. Furthermore, a denoising network is introduced as a precursor to the classification network to further eliminate noise from the spectrum. Finally, self-supervised contrastive learning is employed, utilizing a large number of unlabeled sound samples for pretraining to enhance small-sample generalization and noise robustness. Experimental results demonstrate that the proposed method achieves a 10.6% improvement in accuracy compared to the baseline method, while also exhibiting satisfactory performance in inference time and memory usage.
- 21 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-12
Edge Intelligent Visual Monitoring of Substations Based on Lightweight Object Detection Model
Edge intelligent visual monitoring of substations performs intelligent detection on collected images at edge intelligent perception terminals, representing a novel substation intelligent monitoring paradigm distinct from the cloud computing mode. However, edge computing devices have limited computing power and memory. To enhance the efficiency and feasibility of edge-side object detection algorithms, this paper proposes a lightweight detection algorithm for edge intelligent inspection of substations. Firstly, YOLOv6 is adopted as the baseline model and QARepVGG is introduced to yield QAYOLOv6, making the model conducive to quantization. Secondly, group channel pruning (GCP) and efficient layer pruning (ELP) are proposed to respectively prune the channels and layers of the model, reducing redundant parameters. Finally, quantization-aware training (QAT) is employed to quantize the model to INT8, further reducing computational cost and boosting detection speed. Knowledge distillation is utilized in model training and fine-tuning processes to enhance model accuracy. Experimental results show that on the visible light defect dataset of substations, we achieve a reduction in model computational cost by 63% through pruning and attain over a fivefold acceleration of the model on edge platforms after INT8 quantization, with minimal loss in accuracy.
- 21 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-12
Edge Intelligent Visual Monitoring of Substations Based on Lightweight Object Detection Model
Edge intelligent visual monitoring of substations performs intelligent detection on collected images at edge intelligent perception terminals, representing a novel substation intelligent monitoring paradigm distinct from the cloud computing mode. However, edge computing devices have limited computing power and memory. To enhance the efficiency and feasibility of edge-side object detection algorithms, this paper proposes a lightweight detection algorithm for edge intelligent inspection of substations. Firstly, YOLOv6 is adopted as the baseline model and QARepVGG is introduced to yield QAYOLOv6, making the model conducive to quantization. Secondly, group channel pruning (GCP) and efficient layer pruning (ELP) are proposed to respectively prune the channels and layers of the model, reducing redundant parameters. Finally, quantization-aware training (QAT) is employed to quantize the model to INT8, further reducing computational cost and boosting detection speed. Knowledge distillation is utilized in model training and fine-tuning processes to enhance model accuracy. Experimental results show that on the visible light defect dataset of substations, we achieve a reduction in model computational cost by 63% through pruning and attain over a fivefold acceleration of the model on edge platforms after INT8 quantization, with minimal loss in accuracy.
- Top Cited Papers
- 14 September 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Power Delivery
- Vol. 38 (6) , 4432-4442
CSSAdet: Real-Time End-to-End Small Object Detection for Power Transmission Line Inspection
Ensuring power transmission line reliability is crucial for power system stability. Mechanical joints, critical components of transmission lines, are susceptible to pin defects, leading to severe consequences. We propose an end-to-end lightweight defect detection method, Cross-Scale Spatial Attention Detector (CSSAdet), for accurately identifying pin defects in mechanical joints. CSSAdet integrates spatial and cross-scale attention mechanisms, enhancing feature representation and improving detection accuracy and recall. Our detection pipeline involves two stages: detecting mechanical joints in UAV-acquired images and identifying pin statuses within detected joints, using a single CSSAdet model. We evaluated CSSAdet on a large dataset of transmission line images, comparing it with existing methods. CSSAdet-2 achieves an average precision (AP) of 77.1% and a recall of 91.4% for joint detection, and an AP of 77.2% and a recall of 92.1% for pin status recognition, with a processing speed of 116.8 FPS. CSSAdet demonstrates superior performance in detecting mechanical joint defects with high efficiency, providing a valuable tool for maintaining and monitoring transmission line integrity and improving power system reliability and safety.
- 14 September 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Power Delivery
- Vol. 38 (6) , 4432-4442
Top Cited PapersCSSAdet: Real-Time End-to-End Small Object Detection for Power Transmission Line Inspection
Ensuring power transmission line reliability is crucial for power system stability. Mechanical joints, critical components of transmission lines, are susceptible to pin defects, leading to severe consequences. We propose an end-to-end lightweight defect detection method, Cross-Scale Spatial Attention Detector (CSSAdet), for accurately identifying pin defects in mechanical joints. CSSAdet integrates spatial and cross-scale attention mechanisms, enhancing feature representation and improving detection accuracy and recall. Our detection pipeline involves two stages: detecting mechanical joints in UAV-acquired images and identifying pin statuses within detected joints, using a single CSSAdet model. We evaluated CSSAdet on a large dataset of transmission line images, comparing it with existing methods. CSSAdet-2 achieves an average precision (AP) of 77.1% and a recall of 91.4% for joint detection, and an AP of 77.2% and a recall of 92.1% for pin status recognition, with a processing speed of 116.8 FPS. CSSAdet demonstrates superior performance in detecting mechanical joint defects with high efficiency, providing a valuable tool for maintaining and monitoring transmission line integrity and improving power system reliability and safety.
- Top Cited Papers
- 28 February 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 71 (00189456) , 1-12
A Fast and Accurate Method of Power Line Intelligent Inspection Based on Edge Computing
The combination of power intelligent inspection and edge computing plays a crucial role in the construction of power Internet of Things and transparent power grid. However, as a consequence of the low computing power of edge devices, the detection speed of the model running on edge devices tends to be slow. Accordingly, a fast and accurate method of power line edge intelligent inspection is proposed in this article. First, for key component detection of power lines, RepVGG, diverse branch block (DBB), efficient channel attention (ECA), and improved spatial pyramid pooling (SPP) are utilized to improve YOLOv5, giving rise to RepYOLO. Additionally, to solve the problem of low accuracy in pin defect detection, a two-stage cascaded method is proposed. The first stage localizes various connection fittings in the input image, and the second stage recognizes normal pins and missing pins in these connection fittings. Finally, C++ language combined with TensorRT is employed to optimize and accelerate the model on the NVIDIA Jetson Xavier NX embedded platform, fulfilling efficient power line edge intelligent inspection. Experimental results show that the TensorRT optimized RepYOLO algorithm is 4 times the inference speed of YOLOv5 with a 1.2% increase in accuracy. Moreover, TenosrRT optimized two-stage cascaded method can achieve accurate and real-time pin defect detection on edge devices.
- 28 February 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 71 (00189456) , 1-12
Top Cited PapersA Fast and Accurate Method of Power Line Intelligent Inspection Based on Edge Computing
The combination of power intelligent inspection and edge computing plays a crucial role in the construction of power Internet of Things and transparent power grid. However, as a consequence of the low computing power of edge devices, the detection speed of the model running on edge devices tends to be slow. Accordingly, a fast and accurate method of power line edge intelligent inspection is proposed in this article. First, for key component detection of power lines, RepVGG, diverse branch block (DBB), efficient channel attention (ECA), and improved spatial pyramid pooling (SPP) are utilized to improve YOLOv5, giving rise to RepYOLO. Additionally, to solve the problem of low accuracy in pin defect detection, a two-stage cascaded method is proposed. The first stage localizes various connection fittings in the input image, and the second stage recognizes normal pins and missing pins in these connection fittings. Finally, C++ language combined with TensorRT is employed to optimize and accelerate the model on the NVIDIA Jetson Xavier NX embedded platform, fulfilling efficient power line edge intelligent inspection. Experimental results show that the TensorRT optimized RepYOLO algorithm is 4 times the inference speed of YOLOv5 with a 1.2% increase in accuracy. Moreover, TenosrRT optimized two-stage cascaded method can achieve accurate and real-time pin defect detection on edge devices.
- 5 December 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 241-246
An Energy-Efficient Low-Latency 3D-CNN Accelerator Leveraging Temporal Locality, Full Zero-Skipping, and Hierarchical Load Balance
Three-dimensional convolutional neural network (3D-CNN) has demonstrated outstanding classification performance in video recognition compared to two-dimensional CNN (2D-CNN), since 3D-CNN not only learns the spatial features of each frame, but also learns the temporal features across all frames. However, 3D-CNN suffers from intensive computation and data movement. To solve these issues, an energy-efficient low-latency 3D-CNN accelerator is proposed. Temporal locality and small differential value dropout are used to increase the sparsity of activation. Furthermore, to fully utilize the sparsity of weight and activation, a full zero-skipping convolutional microarchitecture is proposed. A hierarchical load-balancing scheme is also introduced to improve resource utilization. With the proposed techniques, a 3D-CNN accelerator is designed in a 55-nm low-power CMOS technology, bringing in up to 9.89× speedup compared to the baseline implementation. Benchmarked with C3D, the proposed accelerator achieves an energy efficiency of 4.66 TOPS/W at 100 MHz and 1.08 V supply voltage.
- 5 December 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 241-246
An Energy-Efficient Low-Latency 3D-CNN Accelerator Leveraging Temporal Locality, Full Zero-Skipping, and Hierarchical Load Balance
Three-dimensional convolutional neural network (3D-CNN) has demonstrated outstanding classification performance in video recognition compared to two-dimensional CNN (2D-CNN), since 3D-CNN not only learns the spatial features of each frame, but also learns the temporal features across all frames. However, 3D-CNN suffers from intensive computation and data movement. To solve these issues, an energy-efficient low-latency 3D-CNN accelerator is proposed. Temporal locality and small differential value dropout are used to increase the sparsity of activation. Furthermore, to fully utilize the sparsity of weight and activation, a full zero-skipping convolutional microarchitecture is proposed. A hierarchical load-balancing scheme is also introduced to improve resource utilization. With the proposed techniques, a 3D-CNN accelerator is designed in a 55-nm low-power CMOS technology, bringing in up to 9.89× speedup compared to the baseline implementation. Benchmarked with C3D, the proposed accelerator achieves an energy efficiency of 4.66 TOPS/W at 100 MHz and 1.08 V supply voltage.
- Open Access
Trustworthy Online Marketplace Experimentation with Budget-split Design
Online experimentation, also known as A/B testing, is the gold standard for measuring product impacts and making business decisions in the tech industry. The validity and utility of experiments, however, hinge on unbiasedness and sufficient power. In two-sided online marketplaces, both requirements are called into question. The Bernoulli randomized experiments are biased because treatment units interfere with control units through market competition and violate the "stable unit treatment value assumption"(SUTVA). The experimental power on at least one side of the market is often insufficient because of disparate sample sizes on the two sides. Despite the important of online marketplaces to the online economy and the crucial role experimentation plays in product improvement, there lacks an effective and practical solution to the bias and low power problems in marketplace experimentation. Our paper fills this gap by proposing an experimental design that is unbiased in any marketplace where buyers have a defined budget, which could be finite or infinite. We show that it is more powerful than all other unbiased designs in literature. We then provide generalizable system architecture for deploying this design to online marketplaces. Finally, we confirm our findings with empirical performance from experiments run in two real-world online marketplaces.
Open AccessTrustworthy Online Marketplace Experimentation with Budget-split Design
Online experimentation, also known as A/B testing, is the gold standard for measuring product impacts and making business decisions in the tech industry. The validity and utility of experiments, however, hinge on unbiasedness and sufficient power. In two-sided online marketplaces, both requirements are called into question. The Bernoulli randomized experiments are biased because treatment units interfere with control units through market competition and violate the "stable unit treatment value assumption"(SUTVA). The experimental power on at least one side of the market is often insufficient because of disparate sample sizes on the two sides. Despite the important of online marketplaces to the online economy and the crucial role experimentation plays in product improvement, there lacks an effective and practical solution to the bias and low power problems in marketplace experimentation. Our paper fills this gap by proposing an experimental design that is unbiased in any marketplace where buyers have a defined budget, which could be finite or infinite. We show that it is more powerful than all other unbiased designs in literature. We then provide generalizable system architecture for deploying this design to online marketplaces. Finally, we confirm our findings with empirical performance from experiments run in two real-world online marketplaces.