A5048629905- Advanced Neural Network Applications
- Domain Adaptation and Few-Shot Learning
- Multimodal Machine Learning Applications
- Visual Attention and Saliency Detection
- Adversarial Robustness in Machine Learning
- Generative Adversarial Networks and Image Synthesis
- Advanced Image and Video Retrieval Techniques
- Anomaly Detection Techniques and Applications
- Advanced Memory and Neural Computing
- Image Enhancement Techniques
- Handwritten Text Recognition Techniques
- Machine Learning and ELM
- Ferroelectric and Negative Capacitance Devices
- Image Processing and 3D Reconstruction
- Sparse and Compressive Sensing Techniques
- Advanced Photocatalysis Techniques
- Video Surveillance and Tracking Methods
- Catalytic Processes in Materials Science
- Copper-based nanomaterials and applications
- CCD and CMOS Imaging Sensors
- Physical Unclonable Functions (PUFs) and Hardware Security
- Autonomous Vehicle Technology and Safety
- Stochastic Gradient Optimization Techniques
- Olfactory and Sensory Function Studies
- Hand Gesture Recognition Systems
State Grid Corporation of China (China)
2022-2024
Soochow University
2023-2024
North China Electric Power University
2024
China Electric Power Research Institute
2024
Duke University
2017-2020
Alibaba Group (United States)
2020
University of Pittsburgh
2014-2017
Fujitsu (China)
2012-2014
Beijing University of Technology
2009-2011
Huazhong Agricultural University
2010
High demand for computation resources severely hinders deployment of large-scale Deep Neural Networks (DNN) in resource constrained devices. In this work, we propose a Structured Sparsity Learning (SSL) method to regularize the structures (i.e., filters, channels, filter shapes, and layer depth) DNNs. SSL can: (1) learn compact structure from bigger DNN reduce cost; (2) obtain hardware-friendly structured sparsity efficiently accelerate DNNs evaluation. Experimental results show that...
High network communication cost for synchronizing gradients and parameters is the well-known bottleneck of distributed training. In this work, we propose TernGrad that uses ternary to accelerate deep learning in data parallelism. Our approach requires only three numerical levels {-1,0,1}, which can aggressively reduce time. We mathematically prove convergence under assumption a bound on gradients. Guided by bound, layer-wise ternarizing gradient clipping improve its convergence. experiments...
In this paper, a new visual saliency detection method is proposed based on the spatially weighted dissimilarity. We measured by integrating three elements as follows: dissimilarities between image patches, which were evaluated in reduced dimensional space, spatial distance patches and central bias. The inversely corresponding distance. A weighting mechanism, indicating bias for human fixations to center of image, was employed. principal component analysis (PCA) dimension reducing used our...
Very large-scale Deep Neural Networks (DNNs) have achieved remarkable successes in a large variety of computer vision tasks. However, the high computation intensity DNNs makes it challenging to deploy these models on resource-limited systems. Some studies used low-rank approaches that approximate filters by basis accelerate testing. Those works directly decomposed pre-trained Low-Rank Approximations (LRA). How train toward lower-rank space for more efficient DNNs, however, remains as an open...
Deep learning methods have recently achieved impressive performance in the area of visual recognition and speech recognition. In this paper, we propose a hand- writing method based on relaxation convolutional neural network (R-CNN) alternately trained (ATR-CNN). Previous regularize CNN at full-connected layer or spatial-pooling layer, however, focus layer. The convolution adopted our R-CNN, unlike traditional does not require neurons within feature map to share same kernel, endowing with...
Recurrent neural network (RNN) based language model (RNNLM) is a biologically inspired for natural processing. It records the historical information through additional recurrent connections and therefore very effective in capturing semantics of sentences. However, use RNNLM has been greatly hindered high computation cost training. This work presents an FPGA implementation framework training acceleration. At architectural level, we improve parallelism RNN scheme reduce computing resource...
Improving the solar-to-thermal energy conversion efficiency of photothermal nanomaterials at no expense other physicochemical properties, e.g., catalytic reactivity metal nanoparticles, is highly desired for diverse applications but remains a big challenge. Herein, synergistic strategy developed enhanced by greenhouse-like plasmonic superstructure 4 nm cobalt nanoparticles while maintaining their intrinsic reactivity. The silica shell plays key role in retaining superstructures efficient use...
Deep Neural Networks (DNNs) are pervasively used in a significant number of applications and platforms. To enhance the execution efficiency large-scale DNNs, previous attempts focus mainly on client-server paradigms, relying powerful external infrastructure, or model compression, with complicated pre-processing phases. Though effective, these methods overlook optimization DNNs distributed mobile devices. In this work, we design implement MeDNN, local computing system enhanced partitioning...
MXene materials have found emerging applications as catalysts for chemical reactions due to their intriguing physical and applications. In particular, broad light response strong photothermal conversion capabilities are likely render MXenes promising candidates catalysis, which is drawing increasing attention in both academic research industrial satisfy all three criteria of a desirable catalyst: absorption, effective heat management, versatile surface reactivity. However, specific...
Recently, DNN model compression based on network architecture design, e.g., SqueezeNet, attracted a lot attention. No accuracy drop image classification is observed these extremely compact networks, compared to well-known models. An emerging question, however, whether techniques hurt DNNs learning ability other than classifying images single dataset. Our preliminary experiment shows that methods could degrade domain adaptation (DA) ability, though the performance preserved. Therefore, we...
Incorporating phosphorus (P) into the active metals of a catalyst is an effective strategy to enhance catalytic performance. However, mechanisms underlying influence introduced species on performance remain largely unknown. Herein, we observe pronounced shift in product selectivity CO2 hydrogenation from CH4 CO upon introducing P Ru/SiO2 catalysts. This alteration attributed role as "fence" hindering migration H species. The adsorbed CO, key intermediate for methanation, preferentially...
Deep Neural Networks (DNNs) are pervasively used in a significant number of applications and platforms. To enhance the execution efficiency large-scale DNNs, previous attempts focus mainly on client-server paradigms, relying powerful external infrastructure, or model compression, with complicated pre-processing phases. Though effective, these methods overlook optimization DNNs distributed mobile devices. In this work, we design implement MeDNN, local computing system enhanced partitioning...
In Deep Learning, Stochastic Gradient Descent (SGD) is usually selected as a training method because of its efficiency; however, recently, problem in SGD gains research interest: sharp minima Neural Networks (DNNs) have poor generalization; especially, large-batch tends to converge minima. It becomes an open question whether escaping can improve the generalization. To answer this question, we propose SmoothOut framework smooth out DNNs and thereby nutshell, perturbs multiple copies DNN by...
IBM TrueNorth chip uses digital spikes to perform neuromorphic computing and achieves ultrahigh execution parallelism power efficiency. However, in chip, low quantization resolution of the synaptic weights significantly limits inference (e.g., classification) accuracy deployed neural network model. Existing workaround, i.e., averaging results over multiple copies instantiated spatial temporal domains, rapidly exhausts hardware resources slows down computation. In this work, we propose a...
Some recent work revealed that deep neural networks (DNNs) are vulnerable to so-called adversarial attacks where input examples intentionally perturbed fool DNNs. In this work, we revisit the DNN training process includes into dataset so as improve DNN's resilience attacks, namely, training. Our experiments show different strengths, i.e., perturbation levels of examples, have working ranges resist attacks. Based on observation, propose a multi-strength method (MAT) combines with strengths...
As a large-scale commercial spiking-based neuromorphic computing platform, IBM TrueNorth processor received tremendous attentions in society. However, one of the known issues design is limited precision synaptic weights. The current workaround running multiple neural network copies which average value each weight close to that original network. We theoretically analyze impacts low data chip on inference accuracy, core occupation, and performance, present probability-biased learning method...
Road scene understanding and semantic segmentation is an on-going issue for computer vision. A precise can help a machine learning model understand the real world more accurately. In addition, well-designed efficient be used on source limited devices. The authors aim to implement high-level, in embedded device with finite power resources. Toward this goal, propose ApesNet, pixel-wise network which understands road scenes near real-time has achieved promising accuracy. key findings authors'...
Here we present an effective strategy to achieve strongly enhanced catalytic activity of platinum–copper bimetallic clusters through augmented plasmonic photochemical effects aggregated nanostructure.
The exploitation of hierarchical carbon nanocages with superior light-to-heat conversion efficiency, together their distinct structural, morphological, and electronic properties, in photothermal applications could provide effective solutions to long-standing challenges diverse areas. Here, we demonstrate the discovery pristine nitrogen-doped as supports for highly loaded, small-sized Ru particles toward enhanced CO2 catalysis. A record CO production rate 3.1 mol·gRu-1·h-1 above 90%...
Very large-scale Deep Neural Networks (DNNs) have achieved remarkable successes in a large variety of computer vision tasks. However, the high computation intensity DNNs makes it challenging to deploy these models on resource-limited systems. Some studies used low-rank approaches that approximate filters by basis accelerate testing. Those works directly decomposed pre-trained Low-Rank Approximations (LRA). How train toward lower-rank space for more efficient DNNs, however, remains as an open...
In this paper, we propose a vehicle detection method based on AdaBoost. We focus the of front-view car and bus with occlusions highway. Samples different occlusion situations are selected into training set. By using basic rotated Haar-like features extracted from samples in set, train an AdaBoost-based cascade detector. The performance tests static images short time videos show that (1) our approach detects cars more effectively than buses (2) real-time video proceeds at 30 frames per second.