Browsing by Author "Tseng, Fan-Hsun"

Filter results by typing the first few letters
Now showing 1 - 2 of 2
  • No Thumbnail Available
    Item
    Intelligent data cache based on content popularity and user location for Content Centric Networks
    (2019-12-26) Wu, Hsin-Te; Cho, Hsin-Hung; Wang, Sheng-Jie; Tseng, Fan-Hsun
    Abstract Content cache as well as data cache is vital to Content Centric Network (CCN). A sophisticated cache scheme is necessary but unsatisfied currently. Existing content cache scheme wastes router’s cache capacity due to redundant replica data in CCN routers. The paper presents an intelligent data cache scheme, viz content popularity and user location (CPUL) scheme. It tackles the cache problem of CCN routers for pursuing better hit rate and storage utilization. The proposed CPUL scheme not only considers the location where user sends request but also classifies data into popular and normal content with correspond to different cache policies. Simulation results showed that the CPUL scheme yields the highest cache hit rate and the lowest total size of cache data with compared to the original cache scheme in CCN and the Most Popular Content (MPC) scheme. The CPUL scheme is superior to both compared schemes in terms of around 8% to 13% higher hit rate and around 4% to 16% lower cache size. In addition, the CPUL scheme achieves more than 20% and 10% higher cache utilization when the released cache size increases and the categories of requested data increases, respectively.
  • No Thumbnail Available
    Item
    MiniNet：密集擠壓之深度可分離卷積於圖像分類
    (2020) 高汎宜; Kao, Fan-Yi
    近年來，人工智慧的發展蒸蒸日上，自卷積神經網路被提出之後，深度學習開始蓬勃發展，研究學者們紛紛提出更為優化與創新的技術，相較於其它科學領域，深度學習領域的研究採完全開放的方式進行，Google團隊提出TensorFlow開放原始碼函式庫，並在TensorFlow核心庫中支援高階深度學習框架的Keras，幫助開發者在Keras中建立及訓練深度學習模型。未來人工智慧的應用將無所不在，為普及自動駕駛、無人商店、智慧城市等應用，如何在有限的硬體設備中，提供一個運算快速且低計算成本的神經網路模型已成為一個很重要的研究議題。 本論文基於MobileNet架構，加入密集連接技術與擠壓式的SENet模型，提出一個密集擠壓之深度可分離卷積架構，並將此模型命名為MiniNet。本論文在實驗環境中，使用Keras進行MiniNet的建立與訓練，在五種不同的資料集中，與三個現有的卷積神經網路架構進行比較，實驗結果顯示，本論文提出之MiniNet架構能夠明顯地使用更少的計算參數量並有效地縮短訓練時間，尤其在資料集之種類與資料量較少時，本論文提出之MiniNet架構更能優於現有架構達到最高的準確率。