In plain words
You step outside and the grass is wet. Did it rain, or did someone run the sprinkler? You weigh both, and if you then notice the road is dry, you suddenly blame the sprinkler. That kind of reasoning is what a Bayesian network does on a computer.
It is drawn as circles for things that may or may not be true (rain, sprinkler on, wet grass) and arrows for what tends to cause what. Each circle holds numbers: how likely it is, given what points into it.
When you tell the network one fact, it updates every other belief in the map, forwards from causes to effects and backwards from effects to likely causes. It is named after Thomas Bayes, an 18th-century Englishman whose rule for updating beliefs with evidence is at its heart. Unlike many neural networks, you can read the map and see why it reached its answer.
Everyday example
The scooter that won't start
Your scooter won't start. Usually that means an empty tank or a weak battery. You glance at the fuel gauge: half full. Now the battery becomes the prime suspect, without you ever testing it. You also remember the headlight was left on last night, which makes it more likely still. Each clue nudges your beliefs, exactly as the network nudges its numbers.
Where the example breaks: You know scooters from experience. A Bayesian network only knows the arrows and numbers it was given or learned from data; if a real cause is missing from the map, it can never suspect it.
How it actually works
Formally, a Bayesian network is a directed graph with no loops, where each node is a variable and each arrow a direct dependence, plus a table of probabilities at every node.
Draw the structure
Arrows go from causes to effects: Rain → Wet grass, Sprinkler → Wet grass. A node with no arrow between it and another is assumed not to directly affect it. The structure can come from experts, from data, or both.
Fill in probability tables
Each node gets a table of probabilities for every combination of its parents. Wet grass might be 90% likely if it rained, 80% if the sprinkler ran, 5% if neither. Because each node only looks at its parents, the whole model needs far fewer numbers than listing every possible combination.
Add evidence and update
Mark a node as observed, say wet grass is true. Using Bayes' rule, the network recalculates the others. Reasoning from effect back to cause is called diagnosis, and it is why these networks suit medical diagnosis: given symptoms, estimate the chance of each disease.
Explaining away
If wet grass is seen and then rain is confirmed, belief in the sprinkler drops again, because rain already explains the evidence. Handling competing causes this way is one of the network's strengths.
From correlation to cause
Judea Pearl, who named and developed these networks, later extended them to reason about interventions: what happens if we switch the sprinkler on, rather than just observe it. That causal side of his work matters for explainability: knowing not just what goes together, but what causes what.
- Named by
- Judea Pearl, UCLA, in an April 1985 report, "Bayesian Networks: A Model of Self-Activated Memory for Evidential Reasoning"
- Recognition
- Pearl received the 2011 ACM Turing Award for a calculus for probabilistic and causal reasoning
- Named after
- Thomas Bayes, whose essay on probability was published in 1763, two years after his death
- Also called
- Bayes net, belief network
Questions people ask
What is a Bayesian network in simple words?
A diagram of causes and effects with probabilities attached. When you tell it one thing is true, it works out how likely everything else in the diagram has become, much as a detective revises suspicions with each clue.
What is the difference between a Bayesian network and a neural network?
A Bayesian network is a readable map: each circle is a real-world thing and each arrow a stated link. A neural network is layers of numbers learned from data, powerful but hard to inspect. They solve different kinds of problems.
What is a Bayesian network used for?
Reasoning under uncertainty where causes matter. The classic example is estimating diseases from symptoms. Pearl's ideas spread into machine learning, robotics, computer vision, language processing and computational biology.
What is Bayes' theorem?
A rule for updating a belief when new evidence arrives: start with how likely something seemed, then adjust by how well it explains what you just saw. Bayesian networks apply this rule across a whole web of connected beliefs.
Who invented Bayesian networks?
Computer scientist Judea Pearl at UCLA coined the name in 1985 and developed the methods, set out in his 1988 book Probabilistic Reasoning in Intelligent Systems. He won the 2011 Turing Award partly for this work.
Was this clear?
1 is confusing, 5 is crystal clear. One tap, no account needed.
Spotted something wrong, or know a better example? Suggest a change