http://chineseinput.net/에서 pinyin(병음)방식으로 중국어를 변환할 수 있습니다.
변환된 중국어를 복사하여 사용하시면 됩니다.
안신영(Shinyoung Ahn) 대한전자공학회 2023 대한전자공학회 학술대회 Vol.2023 No.6
As artificial intelligence technology matures, the competition for developing DNNs with higher accuracy has become increasingly intense. To achieve higher performance, the size of DNNs is increasing, leading to a significant increase in the time and cost of development. In this paper, we propose EDDIS, a distributed DNN training platform that integrates heterogeneous GPU resources to provide high-speed distributed training to support faster DNN training. EDDIS provides a methodology for modifying existing Tensorflow/PyTorch codes to enable distributed training and also offers an asynchronous parameter update method to solve the straggler problem associated with synchronous parameter update methods, thus providing superior distributed training performance.
안신영 ( Shinyoung Ahn ),이유경 ( Yookyung Lee ),박명호 ( Minghao Piao ),변정용 ( Jeongyong Byun ) 한국정보처리학회 2016 한국정보처리학회 학술대회논문집 Vol.23 No.2
생활수준의 향상 및 소비자들의 건강에 대한 관심의 증가로 인해 자신의 건강에 대해서 스스로 결정하고자 하는 요구가 점차 증가하고 있다. IT 와 의료기술의 발달은 이를 가능하게 하였으며 각종 의료정보를 기반으로 하는 질병진단에 대한 연구가 많이 진행되고 있다. 본 논문에서는 국민건강정보 기반 진료과목 예측에 대한 연구를 진행하여 소비자 스스로 진료과목을 선택하는데 도움을 주고자 한다.
Multi-GPU 기반 분산 딥러닝을 위한 효율적인 집합 통신 기법 연구
최현성(Hyeonseong Choi),이재현(Jaehyun Lee),안신영(Shinyoung Ahn) 대한전자공학회 2023 대한전자공학회 학술대회 Vol.2023 No.6
This paper demonstrates a scheme that can efficiently use collective communication, which is mainly used in multi-GPU-based distributed deep learning environments, through experiments in various environments. We propose a novel buckerting function by referring to the Bucketing scheme proposed for efficient communication in the PyTorch DDP (Distributed Data Parallel) module. In addition, we implemented a communication backend that could handle communication requested by PyTorch through NCCL (NVIDIA Collective Communication Library), and evaluated its performance by applying it to an environment consisting of various types of multiple GPUs. The proposed Bucketing scheme we implemented improved performance by up to about 11% compared to PyTorch DDP and up to about 10.8% compared to Horovod when fine-tuning Bert-base-cased models with GLUE MNLI in the same environment.