본문 바로가기 주메뉴 바로가기
22대 대한민국 국회 국회정보길라잡이 국회의원검색
회원가입 로그인
HOME

정보검색

소장정보 검색

국회도서관 홈으로 정보검색 소장정보 검색

결과 내 검색

동의어 포함

고급검색

인명/단체명

저자정보 상세정보
인명/단체명을 입력하세요.

키워드

  • 대표어

  • 외국어

-

목차보기

Title Page

Abstract

Contents

Ⅰ. Introduction 11

Ⅱ. Motivations and Challenges 16

2.1. Challenges 18

Ⅲ. Overview - On-Modem Learning Framework 20

3.1. On-Modem DNN Inference 21

3.2. On-Modem DNN Training 21

Ⅳ. On-Modem Inference 23

4.1. Data Payload Parser 23

4.2. Data Provider 24

4.3. On-Modem DNN Engine 25

Ⅴ. On-Modem Training 26

5.1. Greedy and Diverse Gradient Accumulation (GDGA) 27

5.2. Dynamic Programming-based Back-Propagation (DPBP) 28

5.3. On-Modem Reinforcement Learning (OMRL) 31

Ⅵ. Application Implementation 34

6.1. Application 1: Remaining Wash Time Prediction 34

6.2. Application 2: Unbalance Error (UE) Forecasting 36

Ⅶ. Experiment 38

7.1. End-to-End Application Performance 38

7.2. Data Payload Parser 42

7.3. Greedy and Diverse Gradient Accumulation (GDGA) 43

7.4. Dynamic Programming-based Back-Propagation (DPBP) 45

7.5. On-Modem Reinforcement Learning (OMRL) 46

Ⅷ. Discussions and Limitations 48

Ⅸ. Related Work 50

Ⅹ. Conclusion 52

References 53

List of Figures

Figure 1. A home appliance without on-modem learning (upper) and one equipped with on-modem learning (lower). In the absence of on-modem learning, the modem module functions solely as a commu... 12

Figure 2. (Left) An exemplary composition of LG Electronics home appliance products represented with model numbers. They are shipped with the identical modem software binary. (Right) Examples of... 18

Figure 3. An overview of the on-modem learning framework. The communication modem module on a home appliance, originally designed for blind data transfer without awareness of its details, carries... 20

Figure 4. The Data Payload Parser is a Long Short-Term Memory (LSTM) architecture based on Recurrent Neural Network (RNN). Its primary function is to recurrently... 23

Figure 5. Data Provider manages requests on the necessary data fields from multiple DNN models and delivers them on demand. 24

Figure 6. Greedy and Diverse Gradient Accumulation (GDGA) on the i-th iteration of back-propagation with data sample xi.[이미지참조] 27

Figure 7. The code implementation of Dynamic Programming-based Back-Propagation (DPBP) written in C language. 30

Figure 8. On-Modem Reinforcement Learning (OMRL) empowers a modem to dynamically and effectively update the value function, represented as a DNN model, in... 32

Figure 9. We implement the proposed on-modem learning on a modem module integrated into an LG washing machine, as depicted in the figure. Furthermore, we... 32

Figure 10. A snippet of superset data used for the remaining wash time prediction. 34

Figure 11. The process involves predicting the remaining wash time on the modem, adjusting with user data (data fields) that reflect the run-time status and dynam-... 35

Figure 12. The Unbalance Error (UE) signals that the washing machine is in the process of redistributing the laundry load to achieve balance. This situation could arise due to either... 37

Figure 13. The MAE (Mean Absolute Error), maximum, and minimum absolute error of the remaining wash time prediction in the unit of minute on the test data over different α values in Figure 11. 'function'... 38

Figure 14. The graph depicts the remaining washing time prediction on the test data over time. In this scenario, the on-modem DNN model is continuously updated (re-trained) with run-time data, while the... 39

Figure 15. The test classification accuracy of UE forecasting over different sequence lengths. Here, 'accuracy' denotes the overall accuracy. 'Non-UE accuracy' and 'UE accuracy' denotes the accuracy of... 40

Figure 16. The confusion matrix of the UE forecasting. Here, 'UE' denotes the class of UE, and 'X' denotes the non-UE. The numbers (1), (2), (3), and (4)... 41

Figure 17. The test parsing accuracy of Data Payload Parser over training epochs. 42

Figure 18. The test classification accuracy on MNIST over training epochs: normal learning vs. Greedy and Diverse Gradient Accumulation (GDGA) with top N＝1,000. 43

Figure 19. The test classification accuracy of Greedy and Diverse Gradient Accumulation (GDGA) on MNIST over training epochs with various Top N. 43

Figure 20. The amount of memory space in bytes required by normal training and Greedy and Diverse Gradient Accumulation (GDGA) with different top N settings. 44

Figure 21. The test accuracy comparison between the normal training method and Dynamic Programming-based Back-propagation (DPBP) on MNIST. 45

Figure 22. The running time and run-time memory usage in training the DNN model on MNIST [28] with and without Dynamic Programming-based Back-propagation (DPBP). 45

Figure 23. The gap between the ground-truth (real) remaining wash time and three prediction methods over 600 washing operations (episodes) on six different washing machines. 'Function' denotes the ex-... 46

초록보기

 We introduce the concept of on-modem learning, a novel approach that empowers communication modems embedded in smart home appliances to conduct inference and training of compact-sized DNNs (deep neural networks) using real-time user data. This enables the delivery of personalized intelligent services within the home environment. Recognizing that a significant portion (80-90%) of commercial home appliances lack network connectivity, resulting in idle computing resources within the modem, our proposed on-modem learning transforms a data-oblivious and non-learning modem into an independent self-learning module. This is achieved by leveraging the unused or under-utilized hardware of the modem for learning from user data. To achieve this objective, we introduce an on-modem DNN inference framework, comprising a Data Payload Parser and On-Modem DNN engine. The former identifies unknown raw data, converting it into a series of recognizable data fields, while the latter provides runtime support for DNN execution on the modem. Additionally, we propose an on-modem DNN training framework that facilitates resource-efficient DNN training on the modem through Greedy and Diverse Gradient Accumulation (GDGA), Dynamic Programming-based Back-Propagation (DPBP), and On-Modem Reinforcement Learning (OMRL). These collectively enhance the user experience in home appliances by enabling real-time learning from customer data and home appliance dynamics, utilizing the limited computing resources (i.e., processor and memory) of the modem and unlabeled user data.

To the best of our knowledge, our on-modem learning approach is the first to execute DNN inference and training directly on the modems of commercial home appliances. To validate the effectiveness of on-modem learning, we implement two on-modem learning applications on an LG Electronics washing machine, i.e., remaining wash time prediction and UE (unbalance error) forecasting. Moreover, we collect over six million user data samples from actual LG home appliance customers and apply them to the two application scenarios. The experimental results demonstrate the successful execution of DNN inference and training on modems integrated into commercial home appliances, providing tailored intelligent services to users.

추천서가 (다양한 추천 자료를 만나보세요)

권호기사보기

권호기사 목록 테이블로 기사명, 저자명, 페이지, 원문, 기사목차 순으로 되어있습니다.
기사명 저자명 페이지 원문 기사목차

권호

기사명 원문보기 기사목차
닫기