Knowledge Map

Summary

The knowledge map shows how the wiki’s major areas depend on each other. It is a conceptual route through prerequisites, not just a list of folders: it answers “what should I understand before this topic makes sense?” rather than “where is this file?” Use learning paths when you want an ordered study sequence, and navigation when you know the destination and just need the shortest route.

The dependency graph

Arrows point from a prerequisite to what it enables. Read top to bottom: mathematics and probability sit at the root, core modeling families branch into modality-specific systems, retrieval and generative AI compose several of those, and production and governance sit downstream of everything applied.

flowchart TD
  Math[Mathematical Foundations] --> Prob[Probability and Statistics]
  Prob --> CML[Classical Machine Learning]
  Math --> DL[Deep Learning]
  Prob --> DL
  Data[Data Engineering] --> CML
  Data --> DL
  CML --> Rec[Recommendation Systems]
  CML --> TS[Time-Series Forecasting]
  DL --> NLP[Natural Language Processing]
  DL --> CV[Computer Vision]
  DL --> RL[Reinforcement Learning]
  CV --> Video[Video Understanding]
  NLP --> IR[Information Retrieval and Search]
  DL --> GenAI[Generative AI and Agents]
  IR --> GenAI
  RL --> GenAI
  Rec --> Prod[ML Engineering and MLOps]
  TS --> Prod
  GenAI --> Prod
  Prod --> Eval[Experimentation and Evaluation]
  Prod --> Gov[Responsible AI and Governance]

The layers

Foundations. Mathematical foundations — linear algebra, calculus, and optimization — and probability and statistics are the shared vocabulary for every model that follows. Gradients, distributions, expectation, and matrix decompositions reappear under different names across the whole tree.

Core modeling. Classical machine learning and deep learning are the two trunks. Classical ML dominates tabular and small-data problems; deep learning dominates perception, language, and generation. Most applied areas specialize one of these.

Modality-specific systems. Recommendation systems, time-series forecasting, natural language processing, computer vision, video understanding, and reinforcement learning adapt the core trunks to a data shape and a decision. They share methods but differ in their evaluation and failure modes.

Retrieval and generative AI. Information retrieval and generative AI are compositional: a retrieval-augmented or agentic system reuses embeddings, ranking, language models, and sometimes preference-based reinforcement learning at once.

Production and governance. ML engineering and MLOps, experimentation and evaluation, and responsible AI sit downstream of every applied area, because a model only becomes a system once it is deployed, measured, and governed.

Cross-cutting concerns

Some areas are not a single layer; they cut across the whole map:

• Data engineering feeds every model, so data quality and pipeline correctness bound the accuracy of everything downstream.
• Cloud and distributed systems and software engineering determine whether a method survives contact with real scale and real teams.
• Experimentation and evaluation and responsible AI apply to every applied area, not just the ones that happen to mention them.

Worked example: what a RAG system depends on

A retrieval-augmented generation system looks like one topic but is a composition of many. Reading up the dependency chain, it relies on embeddings and chunking, dense retrieval and hybrid search, retrieval pipelines, RAG evaluation, PII protection, model serving, and cost and latency optimization. The map exists so a reader treats RAG as this dependency set rather than one isolated box.

Practical use

Use this page when you know a destination topic but not its prerequisites: trace the arrows backward until you reach something you already understand, then study forward. Use learning paths when you want that trace pre-packaged as an ordered sequence, and technical answer patterns when the goal is to explain a concept concisely rather than to learn it.