STRATA
Skip to study
XLVII

IV.07 / AI

Bayesian AI

Which predictions change most when a different prior is fitted to the same evidence?

A source-led introduction2 source references1 related experiments

Start with the question.

Bayesian AI makes uncertainty about parameters, models, or predictions explicit through probability distributions. MacKay's work on interpolation connects this framework to regularization and model comparison. The important questions are what is uncertain, what prior assumptions are encoded, and whether the resulting predictions remain calibrated when evaluated on new observations.

Update a simple predictive belief.

Begin with a uniform prior for a coin's probability of heads. After observing three heads and one tail, the corresponding beta distribution becomes Beta(4,2), with a mean of two thirds. Compare this estimate with the raw frequency of three quarters and inspect how the difference shrinks as more observations accumulate.

What would you inspect?

  • Prior and posterior distributions, predictive intervals, and held-out calibration.
  • Sensitivity of a conclusion to reasonable changes in prior assumptions.

Take the next step.

Fit a small predictive example with two stated priors and report where the data dominate the result and where the prior remains influential.

Put part of the question to work.

These models isolate a mechanism you can vary. Their assumptions define the part of this topic they can illuminate.

Try evidence and base rates.

Change one parameter, compare the numerical output, and export the result. The experiment loads when you open it.

Open its full page

Watch the quantities change.

12 seconds · Original model visualization · Silent

Evidence and base rates

Follow a screening result back through the population that produced it.

Read the explanation and method

A hypothetical batch contains 10,000 components. Detection is fixed at 99 percent and false flags at 5 percent.

The defect rate rises from 1 percent toward 20 percent. Orange shows actual defects that are flagged; cream shows sound items that are falsely flagged.

The posterior is true flags divided by all flags. Changing the starting population changes this fraction even when both screening rates stay fixed.

Method. Apply each conditional rate to its own population. True flags equal the defective population multiplied by the detection rate; false flags equal the sound population multiplied by the false-flag rate. Divide true flags by all flags to obtain the posterior probability.

Read the supporting material.

  1. Bayesian interpolation

    David J. C. MacKay / Caltech · Bayesian model comparison and regularization in adaptive models.

  2. An Essay towards solving a Problem in the Doctrine of Chances

    Thomas Bayes, communicated by Richard Price / University of York transcription · The historical essay on inference from observed outcomes; current exercises are explicitly constructed examples.

Follow the comparisons.

Read the neighboring questions.