MLOps 도입 가이드
데이터이쿠/동동구
My Quotes before Read
데브옵스같은 건가?
Key Topics to Remember
PART 1 MLOps 개념과 필요성
CHAPTER 1 왜 지금이고 도전 과제는 무엇인가
p21 ModelOps, AIOps
p25 DataOps
p26 Risk Assessment, Probability, impact, severity, mitigation
p28 intentionality, explainability, Accountability
CHAPTER 2 MLOps 이해관계자들
p34 Accuracy, Precision, Recall
CHAPTER 3 MLOps의 핵심 기능
p44 KPI
p45 PPI, EDA
p48 PDP, Subpopulation
p50 push-to-production
p51 Ground truth, Input drift
p55~56, 133 shadow test(챔피언/챌린저), A/B test
PART 2 MLOps 적용 방법
CHAPTER 4 모델 개발
p67 Generalization Capacity, stateless, deterministic
p69 하이퍼파라미터
p72~73 Derivatives, Enrichemnt, Encoding, Combination, impact coding, modality, attribute, one-hot encoding, embedding, transfer learning
p76 Underfittinh, Overfitting
p78 cross-testing, k-fold corss-validation
p79 prediction, decision, classification
p80~85 Sharpley Value, ICE, fairness, threshold, interating, assumption, randomness, pseudo-random, confusion Matrix, implementation
CHAPTER 5 상용화 준비
p89~90 quantization, pruning, distillation, NLP, inference
p94~95 accuracy, precision, recall, expectation, check, assertion, reproducibility, Auditability
p97 brute force, poisoning attack
p98 correlation, causality, nearest neighbor algorithm
p100~102 processing chain, trigger, exponential, classification, certainty, conformal prediction
CHAPTER 6 상용 배포
p105 아티팩트
p108 sharding
p109 canary release, affinity
p110 ML metrics monitoring
p111 provisioning
CHAPTER 7 모니터링과 피드백 루프
p119 upper bound, lower bound
p122~126 정확도, ROC AUC, 로그 손실, 비용-이익 평가, null hypothesis, alternative hypothesis, feedback loop, identivally distributed, univariate, domain classifier, empirical distribution function
p133, 134 dark launch, perceived risk, randomness, paited sample T-test
p136, 169 experimental design, effect size, cherry-picking
CHAPTER 8 모델 거버넌스
p143 GxP, GCP, Data INtegrity, reliabillity, reliability, attributable, legible, contemporaneous, original, accurate
p144~152 MRM, intentionality, inclusiveness, chum, HITL, steamline
p154 ad-hoc run, RACI(responsible, accountable, consulted, informed)
p157 audit(감사)
PART 3 MLOps 실제 사례
CHAPTER 9 소비자 신용 리스크 관리
p167, 168 prescreen monotonicity
p170 population staility index, characterisic stability index
CHAPTER 10 마케팅 추천 엔진
p173, 174 impression, collaborative filtering, filtering bubble
p176 cannibalization
p179 인솔자 설계(human-in-command design)
p182~ 185 RMSE, freeze, multi-armed bandit, exploration, exploitation, ensemble, counterfactual, orchestrate
CHAPTER 11 소비 예측
p187~190 poert grid, transmission entwork, distribution network, altermator, congestion, horizon, supervisory control and acquisition(SCADA), metering system, synoptic
p191 photovoltaic:PV, exogenous, horizon, consumption, demand
p194~196 statistical extrapolation, quantile, thermal inertia, GAM, ARIMA, exponential smoothing, LSTM, stochasticity, heuristic, anomaly detection
p197 normative disaggregation, mean absolute percentage error(MAPE)
How the Book Changed Me?
코딩이나 툴 사용법 같은 실무적인 내용은 없고 이론적인 내용만 가득하다. 결론은 AI의 품질이 좋으려면 조직간 협력이 가장 중요하다 정도.