http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
Enabling on-board deep learning model training for mobile and embedded systems
JeongGil Ko(고정길) 한국생산제조학회 2021 한국생산제조시스템학회 학술발표대회 논문집 Vol.2021 No.7
On-device deep neural network model training holds the potential to enable a rich set of data privacy-aware and infrastructure-independent personalized mobile applications [1-3]. However, despite recent advancements in mobile and embedded platform's hardware, locally training a complex deep neural network is still a non-trivial task given its resource demands. Such capability can be used to exploit new machine learning concepts which include federated learning and distributed learning. In this work we show that the limited dynamic memory budgets on mobile platforms is the main constraint, and propose meMon as a framework for efficiently managing mobile memory resources for deep neural network training. Specifically, meMon is cored on a heuristic objective function that identifies neural network activations that are i) easy-to-recompute, ii) consumes large memory spaces, and iii) with weak dependency, so that they are prioritized to be removed from the memory when memory resources are scarce. meMon is implemented in CUDA and will be provided as open source for the research community. We evaluate meMon via hardware benchmark-based emulations and real implementations on the Samsung Galaxy S10 (Android 10) and NVIDIA Jetson TX2. Our evaluations show that meMon enables local training for five state-of-the-art deep learning models (BERT, ResNet-152, DenseNet-169, Stacked LSTM, DCGAN) by reducing the required memory by more than five-fold and speeds up the training process by two-fold compared to a baseline approach. We also show that meMon successfully adapts to dynamic memory budget variations, suggesting its robustness in real-world use cases.
에너지 하베스팅 네트워크에서 SWIPT를 위한 저복잡도를 갖는 파워 할당 및 분할 알고리즘
이기송,고정길,Lee, Kisong,Ko, JeongGil 한국정보통신학회 2016 한국정보통신학회논문지 Vol.20 No.5
RF신호로부터 전력을 수집하는 에너지 하베스팅 기술은 센서의 전원 공급 문제를 해결함으로써 네트워크의 수명을 향상시킬 수 있는 방안으로 최근 큰 관심을 받고 있다. 본 논문에서는 무선 정보 및 전력 동시 전송을 위한 효율적인 알고리즘을 제안하고자 한다. 먼저, 에너지 하베스팅 네트워크에서 채널의 probability density function을 이용하여 water-level의 하계값을 찾은 후, 이를 기반으로 파워 할당 해를 도출한다. 또한, 최소 필요 획득 에너지 조건을 효율적으로 만족시켜줄 수 있는 파워 분할 방안을 제안하였다. 시뮬레이션을 통해 제안 방안은 기존 방안에 비해 최소 필요 획득 에너지 조건을 보장하면서 평균 데이터 전송률을 향상시키고, 최적해에 비해서는 10% 미만의 미미한 성능 저하가 있었지만 계산 복잡도를 현저히 줄일 수 있음을 보인다. Recently, energy harvesting, in which energy is collected from RF signals, has been regarded as a promising technology to improve the lifetime of sensors by alleviating the lack of power supply problem. In this paper, we try to propose an efficient algorithm for simultaneous wireless information and power transfer. At first, we find the lower bound of water-level using the probability density function of channel, and derive the solution of power allocation in energy harvesting networks. In addition, we derive an efficient power splitting method for satisfying the minimum required harvested energy constraint. The simulation results confirm that the proposed scheme improves the average data rate while guaranteeing the minimum required harvested energy constraint, compared with the conventional scheme. In addition, the proposed algorithm can reduce the computational complexity remarkably with insignificant performance degradation less than 10%, compared to the optimal solution.
채널 추정 오차가 존재하는 에너지 하베스팅 네트워크에서 SWIPT를 위한 파워 할당 및 분할 알고리즘
이기송,고정길,Lee, Kisong,Ko, JeongGil 한국정보통신학회 2016 한국정보통신학회논문지 Vol.20 No.7
차세대 무선 통신 시스템에서는 RF 에너지 하베스팅 기술을 이용하여 센서의 전원 부족 문제를 해결하고자 한다. 본 논문에서는 채널 추정 오차가 존재하는 에너지 하베스팅 네트워크에서 무선 정보 및 전력 동시 전송을 위한 효율적인 알고리즘을 제안하고자 한다. 먼저, 1차원의 완전 검색을 통해 최적의 채널 추정 주기를 찾은 후, MMSE 채널 추정기를 이용하여 채널을 추정한다. 추정된 채널 값을 기반으로, 최소 필요 획득 에너지 조건을 만족시켜주면서 데이터 전송률을 최대화할 수 있는 파워 할당 및 분할 방안을 제안하였다. 시뮬레이션을 통해 제안 방안은 완벽한 채널 추정을 가정한 최적 방안에 비해 10% 미만의 성능 저하가 있었지만, 기존 방안과 비교할 시에는 데이터 전송률을 20% 이상 향상시킴을 확인 하였다. In the next generation wireless communication systems, an energy harvesting from radio frequency signals is considered as a method to solve the lack of power supply problem for sensors. In this paper, we try to propose an efficient algorithm for simultaneous wireless information and power transfer in energy harvesting networks with channel estimation error. At first, we find an optimal channel training interval using one-dimensional exhaustive search, and estimate a channel using MMSE channel estimator. Based on the estimated channel, we propose a power allocation and splitting algorithm for maximizing the data rate while guaranteeing the minimum required harvested energy constraint, The simulation results confirm that the proposed algorithm has an insignificant performance degradation less than 10%, compared with the optimal scheme which assumes a perfect channel estimation, but it can improve the data rate by more than 20%, compared to the conventional scheme.
이병복,홍상기,이계선,김내수,고정길,Lee, Byung-Bog,Hong, SangGi,Lee, Kyeseon,Kim, Naesoo,Ko, JeongGil The Korea Institute of Information and Commucation 2013 정보와 통신 Vol.30 No.10
By interacting with external wireless sensors, smartphones can gather high-fidelity data on the surrounding environment to develop various environment-aware, personalized applications. In this work we introduce the sensor virtualization module (SVM), which virtualizes external sensors so that smartphone applications can easily utilize a large number of external sensing resources. Implemented on the Android platform, our SVM simplifies the management of external sensors by abstracting them as virtual sensors to provide the capability of resolving conflicting data requests from multiple applications and also allowing sensor data fusion for data from different sensors to create new customized sensors elements. We envision our SVM to open the possibilities of designing novel personalized smartphone applications.