📚School

2023 최컴특 기말고사 정리

date
Jun 21, 2023
slug
2023-rcl-final
author
status
Public
tags
Summarize
summary
2023 봄학기 최신컴퓨터 특강 정리입니다.
type
Post
thumbnail
category
📚School
updatedAt
Jul 6, 2023 05:57 AM

9주차 : 유사 데이터 검색 기술 소개 ( 김종익 교수님 )


빅데이터 검색에 필요한것
  • 복잡한 데이터 타입을 지원해야함
  • 에러, 비선형, 표현 차이를 잘 관용 할 수 있어야함
 
자카드 유사도 =
Edit distance ED = minimum number of edit operations
Filtering with Pigeonhole Principle : threshold..?
 
Optimal Signature Selection
+ c = number of candidates, filtering cost → Trade-off= Well balanced when c =2
어떻게 선택? → DP →
 
Graph Edit Distance ( GED )
  • 정점과 선의 갯수 차이
 
검증 단계에서 어떻게 GED 연산 오버헤드를 줄일까?
  • 안맞는 파티션을 찾는 온라인 파티셔닝 알고리즘을 개발
  • 만약 안맞는 파티션이 threshold보다 높을경우 GED 연산을 피할 수 있다.
 
PRo-compuation-Based FIltering
  • 쿼리 그래프가 데이터 그래프와 유사하다면 쿼리 결과는 데이터 그래프의 한 부분집합이다.
 
Efficient GED computation
  • GED : 그래프 페어 사이에 가능한 모두 vertex를 열거한다
  • 가장 작은 Graph edit 연산을 열거하는것을 찾는다.
 
 
 

11주차 : DSLAB ( 김경섭 교수님 )


인공지능을 기술을 이용한 유전 정보 분석
  • Assembly process
  • Encoding problem in genomes
    • Contig를 Vector로 어떻게 바꿀 것인가?
  • Genomic Signal Processing
  • Sliding Windows
  • K-mer composigion vector
  • AE
  • AE (Encoder + Clustering): params lower?
  • Kalman Filter , RNN
    • Kalman : process noise와 measurement nosie
    • RNN : 신경망
 

11주차 : Edge computing ( 최훈 교수님 )


엣지 컴퓨팅 : 데이터 소스의 근처에서 실행
 
 
Mobile Edge 컴퓨팅
  • 이동 통신망 기방 edge computing
  • 1ms정도밖에 안걸림
  • 서비스 지연 = 코어망 지연 + 액세스망 지연
 
  • MEC 적용시
    • 스마트폰 - 기지국 - MEC 교환국에서
    •  
       
Vehicluar Edge Computing
  • 차량의 리소스를 통해 데이터를 처리
  • 클라우드 컴퓨팅에 비해 지연시간이 짧고
  • 원격 서버나 데이터 센터로 전송하는 데이터량을 줄임
  • 거리에 있는 풍부한 컴퓨팅 리소스를 활용
 

12주차 : 분산컴퓨팅 및 분산학습 ( 양희철 교수님 )


분산 시스템을 추구하는 이유
  1. 데이터의 증가와 연산량 때문에..
  1. 소통 지연율을 줄일 수 있다.
  1. 위협으로부터 유연한 복구
  1. node failure를 막을 수 있다.
 
분산 컴퓨팅 프레임워크
  • 일반적인 분산 컴퓨팅 모델
  • 맵 리듀스 프레임워크
 
Straggler Effect
  • 막힘 현상 : 높은 지연 아웃라이어가 주는 효과
 
Communication overhead
  • shuffle transper가 MapReduce에서 얼마나 시간을 사용하는가
 
 
Trade-off between straggler effect and computation load
  • Stragger effect = Low computation Latency
  • Computation LOad = No starggler effect
 
Distributed learning
  • Motivation
    • 모델이 커지고 있다 : 계산 비용
  • 병렬 연산
    • 데이터 병렬연산
      • 데이터를 GPU N개에 나눠서 학습시킴
      • 병렬 연산 쉽고, 높은 유용성
    • 모델 병렬 연산
      • 모델을 GPU N개에 나눠서 학습시킴
      • 병렬 연산 어렵고, 로드 밸런싱 이슈
  • Algorithms
    • 연합 학습
      • 엣지와 디바이스 사이에서 합동학습
    • 분할 학습
      • 네트워크위에서 합동 추론
      • Peer-to-peer split learning
    • 탈중앙화 학습
 
 

13 주차 : 신뢰 컴퓨팅 ( 장진수 교수님 )


  • Rest, Transit, Use 할때 Protection
  • 앞 두개는 Prevalent data security model
  • 나머지 3개는 Holistic data security model
  • MPC and HE
    • Secure Multiparty computation
    • Homomorophic Encryption
    •  
Trusted Execution Environment
  • Intel SGX
  • AMD SEV
  • ARM TrustZone
 
Intel SGX
  • MEE - ( Processor Key ) - 물리메모리 ( EPC ) - 주소 공간
  • Processor Key - Snooping
 
ARM TrustZone
  • Normal World
  • Secure World
 
SamSung Knox : Realtime kearel Protection
 
Bypass the SEP
  • GrayKey
  • Cellebrite UFED
 

13 주차 : Representation Learning on Heterogeneous Graphs ( 임성수 교수님 )


  • Node Classification
  • Link Prediction
 
  • Community/anomaly detection\
    • partial node label → predict missing node labels
  • Graph classification & regression
    • partial links → predict missing links
 
Representation learning on graphs
Graph embedding
  • nodes, links, … → summarzie position graph structures or properties
 
Encoder-decode framework
  • Encoder : vector → low-dementional space embedding
  • Decoder : vector embedding to node similarity
  • minimize a reconstruction loss
 
Neural message passing
  • given graph G → find node embeddings
  • update each node embedding using its neighborhood
 
  • Parameters sharing
 
Major GNN models
  • GraphSAGE
    • Node sampling
    • Inductive capability
  • GAT
    • Multi-HEAD attention
  • Taxomomy
    • Heteogeneous graphs
 
 
Knowledge graphs
  • the semantics sets concepts and relations between tehem
  • Knowledge graph embedding
    • minimize a reconstruction loss using the adjacency tensor A
  • MetaPaths : Sequence of node types and link types
  • MetaPath instances : node sequence following the schema defined by P
  • Metapath-based neighbors
    • The Set of nodes which connect with node i via metapath P
 
Maor HGNN
  • HAN
    • semantic-level attention using metapath-based neighbors
  • MAGNN
    • metapath istances connecting the node with its metapath-based neighbors
 
MIGTNet
  • Metapath Instance sampling
  • Linear transformation
  • Generate MI-based graphs
  • MI embedding using GATs
    • Node embedding
    • Parameter learning
 

14 주차: Intelligent Networks ( 김기일 교수님 )


IOT : 빠른전송, 전송의 신뢰성, 저전력
연구분야
  • 애드혹 통신
    • 기지국 없이 무선 통신 기능이 탑재된 디바이스들 간의 네트워크 형성
    • 드론간의 군집비행 : 실시간 위치 정보 동기화가 중요
    • 이상 데이터의 실시간 감지를 위한 통신 경로 최적화 연구
      • 링크 품질, 거리 혼잡도 등 네트워크 상황에 따라 경로를 최적화해야함.
  • 인체영역 통신
    • 디지털 헬스케어 - 초근처리 통신
    • TARA
      • 온도가 낮은 노드를 NEXT-HOP으로 선택
      • HOT-SPOT 지역을 만나면 이전 노드로 되돌아간다.
    • 고신뢰, 저전력 통신 기술 연구
      • 개인화된 이동 패턴을 심층강화학습을 통해 학습
  • 차세대 이동통신
    • 5G 초저지연
  • 모델링 & 시뮬레이션
    • 네트워크 시뮬레이션 : NS-3, OPNET
    • AI for Network