Artificial intelligence(AI)algorithms achieve outstanding results in many applicationdomains such as computer vision and natural language processing The performance ofAl models is the outcome of complex and costly mod...Artificial intelligence(AI)algorithms achieve outstanding results in many applicationdomains such as computer vision and natural language processing The performance ofAl models is the outcome of complex and costly model architecture design and trainingprocesses.Hence,it is paramount for model owners to protect their AI models frompiracy-model cloning,illegitimate distribution and use.IP protection mechanisms havebeen applied to Al models,and in particular to deep neural networks,to verify themodel ownership.State-of-the-art AI model ownership protection techniques have beensurveyed.The pros and cons of Al model ownership protection have been reported.The majonity of previous works are focused on watermarking,while more advancedmethods such fingerprinting and attestation are promising but not yet explored indepth.This study has been concluded by discussing possible research directions in thearea.展开更多
Mobile Edge Computing(MEC) is an emerging technology in 5G era which enables the provision of the cloud and IT services within the close proximity of mobile subscribers.It allows the availability of the cloud servers ...Mobile Edge Computing(MEC) is an emerging technology in 5G era which enables the provision of the cloud and IT services within the close proximity of mobile subscribers.It allows the availability of the cloud servers inside or adjacent to the base station.The endto-end latency perceived by the mobile user is therefore reduced with the MEC platform.The context-aware services are able to be served by the application developers by leveraging the real time radio access network information from MEC.The MEC additionally enables the compute intensive applications execution in the resource constraint devices with the collaborative computing involving the cloud servers.This paper presents the architectural description of the MEC platform as well as the key functionalities enabling the above features.The relevant state-of-the-art research efforts are then surveyed.The paper finally discusses and identifies the open research challenges of MEC.展开更多
Graph deep learning models,which incorporate a natural inductive bias for atomic structures,are of immense interest in materials science and chemistry.Here,we introduce the Materials Graph Library(MatGL),an open-sourc...Graph deep learning models,which incorporate a natural inductive bias for atomic structures,are of immense interest in materials science and chemistry.Here,we introduce the Materials Graph Library(MatGL),an open-source graph deep learning library for materials science and chemistry.Built on top of the popular Deep Graph Library(DGL)and Python Materials Genomics(Pymatgen)packages,MatGL is designed to be an extensible“batteries-included”library for developing advanced model architectures for materials property predictions and interatomic potentials.At present,MatGL has efficient implementations for both invariant and equivariant graph deep learning models,including the Materials 3-body Graph Network(M3GNet),MatErials Graph Network(MEGNet),Crystal Hamiltonian Graph Network(CHGNet),TensorNet and SO3Net architectures.MatGL also provides several pretrained foundation potentials(FPs)with coverage of the entire periodic table,and property prediction models for out-of-box usage,benchmarking and fine-tuning.Finally,MatGL integrates with PyTorch Lightning to enable efficient model training.展开更多
In cognitive radio systems,the design of spectrum sensing has to face the challenges of radio sensitivity and wide-band frequency agility. It is difficult for a single cognitive user to achieve timely and accurate wid...In cognitive radio systems,the design of spectrum sensing has to face the challenges of radio sensitivity and wide-band frequency agility. It is difficult for a single cognitive user to achieve timely and accurate wide-band spectrum sensing because of hardware limitations. However,cooperation among cognitive users may provide a way to do so. In this paper,we consider such a cooperative wide-band spectrum sensing problem with each of the cognitive users able to imperfectly sense only a small portion of spectrum at a time. The goal is to maximize the average throughput of the cognitive network,given the primary network's collision probability thresholds in each spectrum sub-band. The solution answers the essential questions:to what extent should each cognitive user cooperate with others and which part of the spectrum should the user choose to sense? An exhaustive search is used to find the optimal solution and a heuristic cooperative sensing algorithm is proposed to simplify the computational com-plexity. Inspired by this optimization problem,two practical cooperative sensing strategies are then presented for the centralized and distributed cognitive network respectively. Simulation results are given to demonstrate the promising performance of our proposed algorithm and strategies.展开更多
Big data processing is becoming a standout part of data center computation. However, latest research has indicated that big data workloads cannot make full use of modern memory systems. We find that the dramatic ineff...Big data processing is becoming a standout part of data center computation. However, latest research has indicated that big data workloads cannot make full use of modern memory systems. We find that the dramatic inefficiency of the big data processing is from the enormous amount of cache misses and stalls of the depended memory accesses. In this paper, we introduce two optimizations to tackle these problems. The first one is the slice-and-merge strategy, which reduces the cache miss rate of the sort procedure. The second optimization is direct-memory-access, which reforms the data structure used in key/value storage. These optimizations are evaluated with both micro-benchmarks and the real-world benchmark HiBench. The results of our micro-benchmarks clearly demonstrate the effectiveness of our optimizations in terms of hardware event counts; and the additional results of HiBench show the 1.21X average speedup on the application-level. Both results illustrate that careful hardware/software co-design will improve the memory efficiency of big data processing. Our work has already been integrated into Intel distribution for Apache Hadoop.展开更多
As a popular kind of stylized face,cartoon faces have rich application scenarios.It is challenging to create personalized 3D cartoon faces directly from 2D real photos.Besides,in order to adapt to more application sce...As a popular kind of stylized face,cartoon faces have rich application scenarios.It is challenging to create personalized 3D cartoon faces directly from 2D real photos.Besides,in order to adapt to more application scenarios,automatic style editing,and animation of cartoon faces is also a crucial problem that is urgently needed to be solved in the industry,but has not yet had a perfect solution.To solve this problem,we first propose“3D face cartoonizer”,which can generate high-quality 3D cartoon faces with texture when fed into 2D facial images.We contribute the first 3D cartoon face hybrid dataset and a new training strategy which first pretrains our network with low-quality triplets in a reconstruction-then-generation manner,and then finetunes it with high-quality triplets in an adversarial manner to fully leverage the hybrid dataset.Besides,we implement style editing for 3D cartoon faces based on k-means,which can be easily achieved without retrain the neural network.In addition,we propose a new cartoon faces'blendshape generation method,and based on this,realize the expression animation of 3D cartoon faces,enabling more practical applications.Our dataset and code will be released for future research.展开更多
基金supported by the European Union Horizon 2020 research and innovation program under CPSoSAware project(grant no.871738)by Science Foundation Ireland,grant no.12/RC/2289-P2,Insight Centre for Data Analytics。
文摘Artificial intelligence(AI)algorithms achieve outstanding results in many applicationdomains such as computer vision and natural language processing The performance ofAl models is the outcome of complex and costly model architecture design and trainingprocesses.Hence,it is paramount for model owners to protect their AI models frompiracy-model cloning,illegitimate distribution and use.IP protection mechanisms havebeen applied to Al models,and in particular to deep neural networks,to verify themodel ownership.State-of-the-art AI model ownership protection techniques have beensurveyed.The pros and cons of Al model ownership protection have been reported.The majonity of previous works are focused on watermarking,while more advancedmethods such fingerprinting and attestation are promising but not yet explored indepth.This study has been concluded by discussing possible research directions in thearea.
文摘Mobile Edge Computing(MEC) is an emerging technology in 5G era which enables the provision of the cloud and IT services within the close proximity of mobile subscribers.It allows the availability of the cloud servers inside or adjacent to the base station.The endto-end latency perceived by the mobile user is therefore reduced with the MEC platform.The context-aware services are able to be served by the application developers by leveraging the real time radio access network information from MEC.The MEC additionally enables the compute intensive applications execution in the resource constraint devices with the collaborative computing involving the cloud servers.This paper presents the architectural description of the MEC platform as well as the key functionalities enabling the above features.The relevant state-of-the-art research efforts are then surveyed.The paper finally discusses and identifies the open research challenges of MEC.
基金intellectually led by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, Materials Sciences and Engineering Division under contract No. DE-AC02-05-CH11231 (Materials Project program KC23MP). This research used resources of the National Energy Research Scientific Computing Center (NERSC), a Department of Energy Office of Science User Facility using NERSC award DOE-ERCAP0026371. T.W.Ko also acknowledges the support of the Eric and Wendy Schmidt AI in Science Postdoctoral Fellowship, a Schmidt Futures program. We also acknowledged AdvanceSoft Corporation for implementing the LAMMPS interface.
文摘Graph deep learning models,which incorporate a natural inductive bias for atomic structures,are of immense interest in materials science and chemistry.Here,we introduce the Materials Graph Library(MatGL),an open-source graph deep learning library for materials science and chemistry.Built on top of the popular Deep Graph Library(DGL)and Python Materials Genomics(Pymatgen)packages,MatGL is designed to be an extensible“batteries-included”library for developing advanced model architectures for materials property predictions and interatomic potentials.At present,MatGL has efficient implementations for both invariant and equivariant graph deep learning models,including the Materials 3-body Graph Network(M3GNet),MatErials Graph Network(MEGNet),Crystal Hamiltonian Graph Network(CHGNet),TensorNet and SO3Net architectures.MatGL also provides several pretrained foundation potentials(FPs)with coverage of the entire periodic table,and property prediction models for out-of-box usage,benchmarking and fine-tuning.Finally,MatGL integrates with PyTorch Lightning to enable efficient model training.
基金supported in part by the National Basic Research Program (973) of China (No. 2009CB320405)the National Natural Science Foundation of China (No. 60972057)the National High-Tech Research and Development Program (863) of China (No. 2007AA 01Z257)
文摘In cognitive radio systems,the design of spectrum sensing has to face the challenges of radio sensitivity and wide-band frequency agility. It is difficult for a single cognitive user to achieve timely and accurate wide-band spectrum sensing because of hardware limitations. However,cooperation among cognitive users may provide a way to do so. In this paper,we consider such a cooperative wide-band spectrum sensing problem with each of the cognitive users able to imperfectly sense only a small portion of spectrum at a time. The goal is to maximize the average throughput of the cognitive network,given the primary network's collision probability thresholds in each spectrum sub-band. The solution answers the essential questions:to what extent should each cognitive user cooperate with others and which part of the spectrum should the user choose to sense? An exhaustive search is used to find the optimal solution and a heuristic cooperative sensing algorithm is proposed to simplify the computational com-plexity. Inspired by this optimization problem,two practical cooperative sensing strategies are then presented for the centralized and distributed cognitive network respectively. Simulation results are given to demonstrate the promising performance of our proposed algorithm and strategies.
文摘Big data processing is becoming a standout part of data center computation. However, latest research has indicated that big data workloads cannot make full use of modern memory systems. We find that the dramatic inefficiency of the big data processing is from the enormous amount of cache misses and stalls of the depended memory accesses. In this paper, we introduce two optimizations to tackle these problems. The first one is the slice-and-merge strategy, which reduces the cache miss rate of the sort procedure. The second optimization is direct-memory-access, which reforms the data structure used in key/value storage. These optimizations are evaluated with both micro-benchmarks and the real-world benchmark HiBench. The results of our micro-benchmarks clearly demonstrate the effectiveness of our optimizations in terms of hardware event counts; and the additional results of HiBench show the 1.21X average speedup on the application-level. Both results illustrate that careful hardware/software co-design will improve the memory efficiency of big data processing. Our work has already been integrated into Intel distribution for Apache Hadoop.
基金supported by the National Key R&D Program of China(No.2018YFA0704000)the Beijing Natural Science Foundation(No.M22024)+2 种基金the National Natural Science Foundation of China(No.62021002)the Key Research and Development Project of Tibet Autonomous Region(No.XZ202101ZY0019G)supported by the Institute for Brain and Cognitive Sciences,BNRist,Tsinghua University,BLBCI,and Beijing Municipal Education Commission.
文摘As a popular kind of stylized face,cartoon faces have rich application scenarios.It is challenging to create personalized 3D cartoon faces directly from 2D real photos.Besides,in order to adapt to more application scenarios,automatic style editing,and animation of cartoon faces is also a crucial problem that is urgently needed to be solved in the industry,but has not yet had a perfect solution.To solve this problem,we first propose“3D face cartoonizer”,which can generate high-quality 3D cartoon faces with texture when fed into 2D facial images.We contribute the first 3D cartoon face hybrid dataset and a new training strategy which first pretrains our network with low-quality triplets in a reconstruction-then-generation manner,and then finetunes it with high-quality triplets in an adversarial manner to fully leverage the hybrid dataset.Besides,we implement style editing for 3D cartoon faces based on k-means,which can be easily achieved without retrain the neural network.In addition,we propose a new cartoon faces'blendshape generation method,and based on this,realize the expression animation of 3D cartoon faces,enabling more practical applications.Our dataset and code will be released for future research.
