Advisor: Prof. An-Yeu Wu

Presentation on theme: "Advisor: Prof. An-Yeu Wu"— Presentation transcript:

1 Advisor: Prof. An-Yeu Wu
CERES Group Advisor: Prof. An-Yeu Wu

2 Background: Fog Computing for Internet of Things
Smarter Planet Internet of Thing (IoT) The network of physical objects embedded with Sensors (perception layer) Communication (network layer) Computing (application layer) Fog computing Proximity to end-users Latency reduction for quality of service Edge analytics/ stream mining To create value in IoT: Intelligent sensing in perception layer Intelligent computing in application layer

3 Background: Fog Computing for Internet of Things
Power gap Higher power density Limited battery World energy crisis Resource constraints Limited computing hardware and memory in (layered) fog nodes High computational complexity algorithm Thermal issues Learning Task Fog Node Intelligent Sensing: High energy efficiency sensing circuits Intelligent Computing: Light and Green machine learning system

4 Example: 基於步行姿態多重資訊之身分辨識
高解析度影像：人類感官 減少傳輸非必要視訊資料 Kinect: skeleton only (for Machine ONLY)! 機器學習導向之資料! Walking patterns are good biometric traits for people identification 結合多重資訊 環境中的粗略影像 (Light Visual Sensor) 人們身上的穿戴式裝置取得的各項數值

5 Resource Constraints in Fog Computing
Goal: 強健(Robust)且輕量(Light)機器學習系統 海量運算於資源受限環境下 低功耗電壓變異導致運算以及資料錯誤 輕量強健機器學習引擎設計與實踐 Solution: Compressive sensing + ML [2] Reduce sensing and transmission overhead Analyzes data directly in the measurement domain Lose some information Degradation in accuracy Aggregation Model

6 Reliability Issues in ML Framework
Voltage over-scaling (VOS) for energy efficiency [12] PVT variation & soft error because of process scaling [13] Noisy data Computational errors in classifier or trainer Numerical stability Robust learning Robust weak classifiers Error-aware aggregation model Computational Errors Data Noise Robust Learning Error-aware aggregation

7 Proposed Aggregation of Robust and Compressed Learning Framework
Parallel and iterative training algorithm Many-core computing system Robust low-power classifier High-reliable Trainer (offline) Variation-aware core-level redundancy scheme CS-based sensor Feature Vectors [3] Classifier