請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92417
標題: | 利用深度變分學習優化由消息理論建立基於損失之任務導向通信的抗干擾表徵 Deep Variational-Enabled Information-Theoretic Representation Learning with Nuisance Invariance to Loss-Based Task-Oriented Communication |
作者: | 朱禹安 Yu-An Chu |
指導教授: | 蘇炫榮 Hsuan-Jung Su |
關鍵字: | 任務導向通信,裝置邊緣協同推理,多任務學習,抗干擾,深度變分推斷,隨機表徵學習,可變長度特徵編碼, Task-oriented communication,Device-edge co-inference,Multi-task learning,Nuisance invariance,Deep variational inference,Stochastic representation learning,Variable-length feature encoding, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 當前,眾多智能服務仰賴深度神經網絡的推斷,然而在行動裝置(如智慧型手機和感測器)上執行它們需要耗費大量資源。此外,某些實際應用牽涉到基於多個子任務的決策,並且需要即時響應。例如,在自動駕駛中,同時處理物體檢測和車道跟蹤。為了使下游任務維持可接受的性能,一種方法是將部分資源密集型運算轉移到靠近資訊來源的強大邊緣伺服器,同時將每個樣本的充分但最小語義表徵從裝置上行傳輸。
根據上述的任務導向通信,我們協同裝置和邊緣,建立了一個通用而適應性強的推理框架,以實現多任務無損預測。與之前僅適用於少數預定義任務的工作不同,此框架對各種潛在應用任務的適用性僅受到其損失類型(如對數損失)的限制,而且對裝置本身而言這些任務可以完全未知。本論文從消息理論角度推導並解釋正規化特定通信約束的目標函數,順帶展示了其與機器學習中多個熱門領域引人入勝的交集。 實作中的任務導向通信系統包括一個通用的圖像發射器,對於動態通道條件和語義內容具有穩健性。針對可變長度的特徵編碼和連續觸發的神經元,我們設計了可微分的資料流和神經網絡模組,以最小化通信開銷。此外,我們修改了變分丟棄層,並修剪標準化後的激活數值,以在類比傳輸過程中引入符號稀疏性。延遲和多任務分類準確度之間的不同權衡證明了所提架構的有效性,也證實了我們的見解。與資料導向的信源通道聯合編碼相比,免去高維重建導致了較低的延遲和減少的計算需求。 隨著機器學習的不斷發展,這個原型有望套用各個提到的研究領域中更進階的技術,為未來在各種實際應用中帶來令人興奮的前景。 Intelligence services inferring from deep neural networks are too resource-intensive to run efficiently on mobile devices like smartphones and sensors. In addition, some real-world applications make decisions based on multiple subtasks and demand real-time responses. For example, object detection and lane tracking should be handled simultaneously in autonomous driving. To maintain acceptable performance for the downstream tasks, one approach is to offload parts of the heavy computation to a powerful edge server located close to the data source, while the sufficient but minimal semantic representation of each sample is uplink transmitted from the device. In terms of such task-oriented communication, this work establishes a general and adaptable device-edge co-inference framework to achieve multi-task lossless predictions. In contrast to previous works that are only suitable for few pre-defined tasks, various potentially applicable tasks are only restricted to their loss types (e.g., log loss) and can be completely unknown for devices in our framework. The regularized objective function under certain communication constraints is derived and interpreted from an information-theoretic view. It also presents a fascinating intersection of multiple fields in machine learning. The designed practical task-oriented communication system involves a generic image transmitter that is robust to dynamic channel conditions and semantic contents. Differentiable data flow and network modules are designed for variable-length feature encodings and consecutive neuron activations to minimize the communication overhead. Besides, variational dropout is modified to prune normalized activations for inducing dimensional sparsity in the analog transmission manner. Experiments with varying trade-offs between latency and multi-task classification accuracies demonstrate the effectiveness of the proposed framework, which corroborates our ideas. Compared to data-oriented joint source-channel coding, high-dimensional reconstructions are unnecessary, which leads to lower latency and decreases computational requirements. As machine learning continues to advance, this prototype is poised to benefit from improved techniques in each mentioned research area, which promises exciting prospects for future implementations in a wide array of applications. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/92417 |
DOI: | 10.6342/NTU202304514 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 電信工程學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-112-1.pdf 此日期後於網路公開 2028-12-14 | 7.06 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。