Skip to content
Road to Intelligence

Concept · Chapter 1: What Is Artificial Intelligence?

Knowledge Representation

Should knowUnderstand15 minDifficulty

Knowledge representation is the problem of writing down what a system knows — objects, categories, relations, defaults — in a form a machine can reason with.

The problem

Rules need something to talk about: a structured description of the world, its categories and relationships.

The solution

Formalisms such as semantic networks, frames and ontologies organize concepts into categories with properties and relations, with defaults that can be inherited or overridden.

The consequence

These ideas live on in knowledge graphs and schemas — but encoding everyday common sense by hand proved enormous, which pushed AI toward learning representations instead.

You should understand first

  1. The Turing Test
  2. Symbolic AI
  3. Knowledge Representation

The idea

Before a system can reason about restaurants, it needs to know what a restaurant is: a place with a menu, staff and bills, where you order and then pay. Symbolic AI developed ways to write this down:

  • Semantic networks — a graph: canary —is-a→ bird —has→ wings.
  • Frames — structured templates with slots and defaults: a bird frame with can_fly: true that a penguin frame overrides.
  • Ontologies — agreed vocabularies of categories and relations, today's knowledge graphs.

The problem it exposed

Everyday reasoning relies on an enormous amount of unstated common sense: water is wet, people can't be in two places at once, a dropped glass may break. Projects that tried to encode it by hand — Cyc began in 1984 — accumulated vast rule bases and still fell short of the breadth humans take for granted.

What to remember

  • Semantic networks: concepts as nodes, relations as labelled edges.
  • Frames: structured records with slots and default values.
  • Ontologies/knowledge graphs: shared vocabularies of entities and relations.
  • Common sense turned out to be vast and mostly unwritten.