❒ 초대의 글
최근 들어 IoT(사물인터넷)은 우리 일상 생활의 다양한 분야에서 적용되어 패러다임을 바꾸고 있습니다. 여기에 인공지능이 결합되면서 헬스케어, 자율 주행, 보안 감시, 사용자 행동 인식, 교통, 공장 로봇, 위험 관리 등 복잡한 문제에 대하여 지능적이면서 효율적인 해결책을 제시할 가능성을 열어보이고 있습니다. IoT 장치는 고사양 컴퓨터로부터 휴대폰, 계산 능력 및 저장 공간이 제약된 저사양 마이크로컨트롤러에 이르기까지 범위가 다양합니다. 그러나, 인공지능 중에서도 일반적으로 대용량의 계산 능력 및 저장 공간을 필요로 하는 딥러닝을 저사양의 IoT 장치에 적용하는 것은 매우 도전적인 문제입니다. 이 문제에 대한 해결 방안으로, 딥러닝을 제약된 IoT 장치에 경량화 및 최적화시키기 위해 다양한 연구가 이루어져 왔습니다. 본 여름학교에서는 지금까지 이루어진 다양한 연구를 모델 압축이라는 소프트웨어 레벨과 하드웨어 가속이라는 하드웨어 레벨에서 다룹니다. 또한 강화 학습의 기본적인 측면과 더불어 얼굴 인식, 생체 응용 및 연합 학습에서의 이슈 및 연구 현황도 살펴봅니다. 더 나아가 인공지능에 기반하여 미래 네트워크를 설계하기 위한 연구도 소개합니다. 본 행사에서는 강연자와 참여자 간에 여름학교 전용 채널을 운영하오니, 국내외 연구자들이 국경, 분야를 초월하여 지속적으로 교류하고 협력하기를 바랍니다.

인공지능및보안연구회 위원장 황성운
멀티미디어연구회 위원장 심정연
컴퓨터소사이어티 회장 황성운
대한전자공학회 회장 공준진

❒ 행사 개요
o 행사명: 2021년도 대한전자공학회 인공지능및보안연구회/ 멀티미디어연구회 합동 여름학교
              - Advances and Challenges of Artificial Intelligence in the Internetof-Things Era
o 일 시 : 2021년 7 월 15일 (목) - 7월 16일 (금)
o 장 소 : 온라인 
o 주 최 : 대한전자공학회 인공지능및보안연구회, 대한전자공학회 멀티미디어연구회
o 주 관 : 대한전자공학회 
o 후 원 : 
    IEEE Seoul Section Sensors Council Chapter
    가천대학교 BK21 FAST인공지능융합센터 
    홍익대학교 BK21 초분산 자율 컴퓨팅 서비스 기술 연구팀
    국민대학교 특수통신융합서비스연구센터 

❒ 운영위원
o 프로그램 위원장 : 황성운 교수 (가천대)
o 프로그램 위원 : 
    Hyung Jin Chang 교수 (University of Birmingham, 영국)
    고병철 교수 (계명대)
    Boon-Yaik Ooi 박사 (Universiti Tunku Abdul Rahman, 말레이시아)
    Andrew Beng-Jin Teoh 교수 (연세대)
    Wai Kong Lee 박사 (가천대)
    박수현 교수 (국민대)
    김병서 교수 (홍익대) 
    안현식 교수 (동명대)
    심정연 교수 (강남대)

❒ 세부 프로그램
* 강의는 모두 영어로 진행됩니다. 
* https://ai-security.github.io/summer-school-2021을 통해 강연자와 참석자간에 다양한 정보 교류, 질문과 응답을 할 수 있습니다. 
* 발표 자료는 온라인으로 제공되며, 별도의 인쇄본은 제공하지 않습니다.
* 프로그램은 주최 측의 사정에 따라 변경될 수 있습니다. 

첫째날: 2021년 7월 15일 목요일

둘째날: 2021년 7월 16일 금요일

❒ 강연요약

강연자	강연 내용
Warren Powell  교수 (Princeton University)	Reinforcement Learning and Stochastic Optimization: A unified framework for sequential decisions Part 1 & 2
	I am going to provide a single, unified framework that covers every sequential decision problem, ranging from pure learning problems (e.g. running experiments), to complex dynamic resource allocation problems, or stochastic search problems. All of the problems reduce to optimizing over “policies” which are methods for making decisions. I will identify four (meta) classes of policies that include every method studied in the academic literature, or used in practice. Each of the four classes of policies will be illustrated in the context of pure learning problems, stochastic search, and a variety of applications in transportation, energy and health.
Cheng-Yaw Low  박사 (연세대)	New Challenges to Face Recognition: Low-Resolution Face Recognition and Periocular Recognition
	Face recognition has continually accomplished significant breakthroughs thanks to the advancement of deep neural networks (DNN) in association with powerful loss functions and the availability of million-scale training datasets. In place of the typical face recognition problem, this talk emphasizes two new challenges to face recognition, specifically low-resolution (LR) face recognition and periocular face recognition (masked face recognition). We will probe into the major problems for training a high-capacity DNN to confront these challenges. In the meantime, some recent works will also be introduced.
김병서  교수 (홍익대)	AI for Information-Centric Networks as a Future Network Technology
	Even though designed for future network Internet technology, Information Centric Networking (ICN) has been researched for wireless communications because it provides connectionless, non-destination-oriented, caching, etc. Particularly, ICN is well-adopted for massive, content-based, heterogeneous IoT networks. Recently, AI-based ICNs are studied to overcome the issues of ICNs such as optimal forwarding & routing, caching strategy, congestion control, producer mobility, etc. In this talk, after a brief review of ICNs and the issues of ICNs in wired & wireless networks, recent researches for AI-based ICNs are introduced.
고병철  교수 (계명대)	Deep Review of Model Compression in Knowledge Distillation Side
	Deep learning algorithms have higher performance as the layers deepen, but have disadvantages in that memory requirements and processing speed increase as they require many parameters. Therefore, recently, AI studies have been attempted to achieve similar performance while reducing the number of deep learning layers and parameters. Among these methods, there are parameter pruning and sharing, low-rank approximation, and teacher-student networks. In this lecture, each deep learning compression technique will be briefly reviewed, and in particular, the teacher-student networks method will be described in focus. In addition, I will introduce cases in which these compression methods are actually applied.
Andrew Beng-Jin Teoh  교수 (연세대)	Biometric Cryptosystem: Progress and Challenge
	The inability of humans to remember and generate strong secrets makes it problematic for people to manage cryptographic keys. To address this problem, biometric cryptosystem has been put forward to enable a user to repeatedly generate a cryptographic key from his/her biometrics while protecting identity theft. Some prominent instances of biometric cryptosystems are Fuzzy Commitment, Fuzzy Vault and Fuzzy Extractor. Despite biometric cryptosystems have made vital contributions by specifying formal security definitions with where the schemes can be analyzed and provably secure, there remains a huge gap between theoretical soundness and practical systems. In this talk, an overview of progress of biometric cryptosystems will be presented. Specifically, design requirements, pitfalls and subtleties that are commonly overlooked in the practical design and assessment of biometric cryptosystems will be highlighted. Finally, a number of possible remedies addressing the challenges in designing practical biometric cryptosystems are discussed.
박수현  교수 (국민대)	Maritime, Underwater IoT and AI-based First-order logic TUM-IoT Digital Twin
	In order to actively respond to changes in the Internet of Service (IoS) market environment of the Internet of Things (IoT) that exists in the terrestrial / underwater / maritime domain, a key technology for creating an intelligent autonomous service based on situational awareness is required. By combining intelligent underwater and terrestrial IoS interworking services with digital twin technology, it is necessary to provide standard technology that can autonomously create optimal services, thereby providing digitalization services at low cost. It is possible to autonomously create new intelligent services by linking IoS of various IoT in different domains (underwater/sea/terrestrial) according to context.
Jin-Chuan See  박사 (Universiti Tunku Abdul Rahman)	Overview of Model Compression and Quantization in Deep Learning
	Image classification using deep learning is a powerful technique and was shown to produce higher accuracy in recent years. However, high accuracy deep learning networks typically contains large network parameter sizes. This presents an issue when it comes to porting onto resource constrained (in terms of energy, availability CPU/GPU and memory sizes) devices. Model compression techniques were proposed to reduce the large network parameter sizes and at the same time retains the networks’ accuracy. In this session, we first give an overview of the techniques used in model compression, followed by an in-depth discussion on one of the techniques, quantization.
Rehmat Ullah  박사 (Queen's University)	Edge Federated Learning: Recent Advances and Open Research Problems
	Google recently introduced the concept of Federated Learning (FL) in 2016, which is a privacy-preserving ML technique in which an ML model is collaboratively learned across several distributed devices (e.g., mobile phones), while all training data is kept on local devices. The FL provides privacy-by-design and is well suited for edge computing applications because it can take advantage of the computation power of edge servers. This talk will discuss distributed ML, with a focus on FL for edge computing systems. This talk will start by giving a quick explanation of FL, how FL solves the data island problem in IoT and state-of-art advances of FL. The edge federated learning applications with open-source platforms, current trends, and recent development will be discussed particularly form the network perspective. Furthermore, open research challenges with potential solutions will be presented.
송진호  교수 (연세대)	Hardware Acceleration and Optimization of Deep Neural Networks
	Nearing the end of Moore’s Law, computing systems are geared towards specialization, i.e., executing workloads using multiple, heterogeneous processing units. With rapid advances in deep learning techniques, deep neural networks (DNNs) have become important workloads for the computer systems. Such trends sparked recent races to develop neural processing units (NPU), specialized processors for DNNs. This talk discusses various recent efforts to design lightweight neural networks and optimization techniques built into hardware accelerators to enhance computational efficiency. In many cases, software-level optimization ideas for DNN acceleration are tightly coupled with hardware-level supports. A review of hardware-side related work will hopefully be helpful to a large audience of the event.

❒ 등록비 안내