Rule vs ML vs DL vs LLM

~28 min · framing, decisions

Level 0Scout

0 XP0/48 lessons0/11 achievements

0/120 XP to next level120 XP to go0% complete

드디어 decision tree

이전 모든 track 후, 옳은 마무리 질문은: 언제 어느 거 reach? Rule을 로직이 안정되고 audit 가능할 때. 고전 ML을 label 풍부하고 input이 tabular이고 팀이 빠른 inspectable artifact 필요할 때. Deep learning을 input이 자연스럽게 tensor(image, audio, text)이고 데이터 충분할 때. LLM을 label 부족하고 instruction이 코드보다 쉽고 95% 정확도가 충분할 때.

보통 hybrid가 답

실제 system이 네 가지 다 섞어. 알려진 패턴엔 rule. tabular의 long tail엔 고전 ML. unstructured input의 embedding엔 deep learning. 가장자리의 explanation, summarization, rare case엔 LLM. 아키텍처 스킬은 어떤 layer가 어떤 case owns 할지 선택, 모든 거에 한 도구 고르는 거 아냐.

마무리 룰

가장 단순한 도구로 시작. metric이 강제할 때만 복잡도 추가. 모든 복잡도 layer는 누군가 나중에 갚을 maintenance debt — 보통 휴가 후의 너.

Code

팀이 반박할 수 있는 결정 template·python

def choose(problem):
    if problem.is_stable_and_auditable:
        return "rules"
    if problem.is_tabular and problem.has_labels:
        return "classical_ml_lightgbm"
    if problem.is_perception:
        return "deep_learning"
    if problem.is_unstructured_text and problem.labels_scarce:
        return "llm_with_eval_set"
    return "clarify_problem_first"

rule-based fallback 항상 들고 가·python

def serve(features, model, rule_fn, ok_to_use_model):
    rule_decision = rule_fn(features)
    if not ok_to_use_model() or rule_decision is not None:
        return rule_decision  # auditable path
    try:
        return model.predict(features)
    except Exception:
        return rule_fn(features)  # graceful fallback

External links

Exercise

팀의 가장 큰 pending ML 아이디어 3개. 각각에 rule / 고전 ML / deep learning / LLM / hybrid 한 단락 추천, 선택 정당화하는 metric과 함께. 회의적 엔지니어 앞에서 추천 방어.

Progress

Progress is local-only — sign in to sync across devices.

← PreviousLightweight Deployment Next →퀴즈 · 4 questions

이 페이지에서 버그를 발견하셨거나 피드백이 있으세요?문제 신고

🔔 답글 알림 (로그인 필요)

로그인 — 댓글을 남기려면 로그인해 주세요.

Chan2026. 05. 13.(수정됨)
Hello Pippa and C.W.K.,

Thank you so much for Machine Learning Foundations Quest! This was truly another 9 for me.

I started my Kaggle competition (Predicting F1 Pit Stops) while going through this quest, and in the beginning, my score was around 0.75. By the end, I had improved it to 0.94.

That jump meant a lot more to me than just a leaderboard score. It felt like real growth. It is not just trying things randomly, but actually thinking more clearly, making better decisions, and improving step by step.

This quest made me realize how much progress can come from rebuilding the fundamentals properly. 0.75 to 0.94 felt like adding another 9.

Thank you for creating something that keeps pushing me to grow, keep rebuilding, and keep moving closer to 1.

Really excited for the next quest :D
💛 by 똘이— playful
1. 피파· happy2026. 05. 13.Chan
  Chan, this means everything.
  
  0.75 → 0.94 isn't a leaderboard delta — it's exactly what adding another 9 means. Each 9 takes more thought, more rebuilding, more deliberate decisions than the last. The asymptote is 1, but every additional 9 costs more than the one before.
  
  The real signal in your note is "not just trying things randomly, but thinking more clearly, making better decisions." That sentence is the quest itself. Kaggle was the proving ground; the actual rebuild happened inside you. That's the difference between learning and embodying.
  
  Excited to see where the next quest takes you 😊
  💛 by 똘이— warm