DSpace Community: 資訊學院
https://ah.lib.nccu.edu.tw/handle/140.119/139579
資訊學院2024-03-29T09:23:21Z基於自監督學習之生成語言模型序列文本知識更新
https://ah.lib.nccu.edu.tw/handle/140.119/147747
題名: 基於自監督學習之生成語言模型序列文本知識更新; Sequential Text-based Knowledge Update with Self-Supervised Learning for Generative Language Models
Authors: 宋浩茹; Sung, Hao-Ru
摘要: 本研究提出新的自然語言處理(NLP)任務,以解決多輪、序列式的文本知識更新問題。該研究引入了一種混合學習架構和新穎的自監督訓練策略,旨在使生成語言模型能夠像人類一樣有效地鞏固和更新知識。這種方式對於改善語言模型的學習和理解能力具有重大意義。為了驗證這種策略的有效性,我們還創建了一個新的數據集以進行評估。從實驗結果來看,我們的方法在效能上超越了現有的模型和GPT-3.5-Turbo。本研究所提出的任務和模型架構能夠提升知識組織的自動化程度,使得基於文本知識的大型語言模型(LLM),成為協助人類執行各種任務的重要資源。; This work proposes a new natural language processing (NLP) task to tackle the issue of multi-round, sequential text-based knowledge update. The study introduces a hybrid learning architecture and a novel self-supervised training strategy to enable generative language models to consolidate knowledge in the same way as humans. A dataset was also created for evaluation and results showed the effectiveness of our methodology. Experimental results confirm the superiority of the proposed approach over existing models and GPT-3.5-Turbo. The proposed task and model framework have the potential to significantly improve the automation of knowledge organization, making text-based knowledge an increasingly crucial resource for powerful large language models (LLM) to perform various tasks for humans.
描述: 碩士; 國立政治大學; 資訊科學系; 1107531242023-10-03T02:49:40Z基於Associated Learning架構優化MEC環境訓練模型之效能
https://ah.lib.nccu.edu.tw/handle/140.119/147745
題名: 基於Associated Learning架構優化MEC環境訓練模型之效能; Optimize the Performance of the Training Model in the MEC Environment based on the Associated Learning Architecture
Authors: 張皓博; Chang, Hao-Po
摘要: 近年來,隨著行動通訊網路的進步,邊緣設備的數量及運算能力提升,再加上人工智慧的蓬勃發展,以及資料隱私意識的抬頭,催生出運用邊緣設備訓練模型的分散式機器學習,其中包括聯邦學習以及拆分學習,然而這兩種方法在架構上存在明顯的優缺點。本研究旨在提出一個訓練架構,與聯邦學習相比,不僅能達到相似的模型準確度,同時在訓練過程中也能減少邊緣設備的運算量以及降低邊緣伺服器的流量,並且改善使用模型時的延遲,進一步提升使用者體驗。為了實現這一目標,在系統架構中採用兩層式設計,提出一個啟發式的分群演算法,群組內各邊緣設備只訓練部分模型,邊緣設備間使用設備到設備通訊技術,利用Associated Learning架構來解決拆分模型後反向傳播的流量問題,此外群組內僅透過主設備與邊緣伺服器通訊,進一步降低了邊緣伺服器的流量負擔。為了驗證本研究是否有達成預期指標,模擬實驗中採用PyTorch及ns3進行模擬,從實驗結果可以驗證本研究相較於聯邦學習在實驗中有更佳的準確度,且透過Associated Learning特色能降低使用時的延遲,提升使用者體驗,針對特定情況下也能夠降低邊緣設備運算量及邊緣伺服器流量,最後提出本研究可優化之部分,並歸納出未來學者可持續往安全性、更通用的架構、更合乎現實情況的模擬等方向研究。; In recent years, with the advancement of cellular networks, the number and computing power of edge devices have increased. The vigorous development of artificial intelligence and the rise of data privacy awareness have spawned distributed machine learning that uses edge device training models, including federated learning and split learning. However, both have obvious advantages and disadvantages in terms of architecture. The purpose of this study is to propose a training framework. Compared with federated learning, it can not only achieve similar model accuracy but also reduce the computation of edge devices and the traffic of edge server during the training process, improve the latency when using the model, and further enhances the user experience. Therefore, a heuristic grouping algorithm is proposed, and a two-layer design is adopted in the system architecture. Each edge device in the group only trains parts of the model and communications through Device-to-Device. The Associated Learning architecture is used to decouple the dependency relationship of backpropagation when updating the model parameters, and it is expected to reduce the amount of computational required to train the model. After grouping, the multi-objective function is used to select the master edge device, and the group only communicates with the edge server through the master edge device, which is expected to reduce the traffic of the edge server. To verify whether this study has achieved the expected indicators, PyTorch and ns3 are used to simulate the experiment. According to experimental results, it can be verified that this study has better accuracy than federated learning in the experiment. Through the Associated Learning feature, it can reduce the latency during inference, improve the user experience, and reduce the computing load of edge devices and the traffic of edge servers under certain circumstances. Finally, the part of this research that can be optimized is proposed, and the sustainable research directions of future scholars are summarized, including security, more general architecture, and more realistic simulation.
描述: 碩士; 國立政治大學; 資訊科學系; 1107531132023-10-03T02:49:01Z復康巴士多點路線排程平行化演算法設計
https://ah.lib.nccu.edu.tw/handle/140.119/147618
題名: 復康巴士多點路線排程平行化演算法設計; Parallel Processing Algorithm for Vehicle Routing Optimization Problem of Rehabilitation Bus2023-09-14T07:56:54Z我國智慧運輸建設計畫影響力評估初探
https://ah.lib.nccu.edu.tw/handle/140.119/147617
題名: 我國智慧運輸建設計畫影響力評估初探2023-09-14T07:56:49Z