Does more data always make the posterior correct?

More relevant data can reduce uncertainty within a model, but incorrect assumptions or biased observations can still produce misleading conclusions.

Is the posterior a single number?

It can be a probability for a particular event, but the full posterior is a distribution over the unknown quantity.

Can it have several peaks?

Yes. Multiple distinct explanations can remain plausible after the evidence is considered.

AI Glossary / Computer vision and multimodal AI / Intermediate

Posterior

Definition

A posterior is a probability distribution updated using observed evidence within a specified model. In Bayesian inference, it combines a prior distribution with a likelihood, describing uncertainty about unknown quantities after taking the data into account.

What it means

Before a measurement, several explanations may be plausible. The prior describes their initial weighting within the model. The likelihood describes how compatible the observed data are with each explanation. The posterior combines those two sources of information. In computer vision, the unknown might be an object's class, a camera pose, a depth map or a clean image underlying a noisy observation. A posterior is not necessarily one best answer; it can describe several possibilities and their relative probabilities. Its reliability depends on the assumptions used to construct it.

How it works

Bayes' rule gives posterior probability as likelihood multiplied by prior probability, divided by a normalizing quantity representing the evidence. For a continuous unknown, the same idea uses probability densities and an integral for normalization. Computation can be exact in simple models or approximated using optimization, sampling or other inference methods.
Different summaries answer different questions. A posterior mean averages over uncertainty, while a maximum a posteriori estimate selects a highest-density or highest-probability explanation under the representation. Credible intervals or regions summarize uncertainty according to the model. A narrow posterior does not guarantee that the model is correct, particularly if the likelihood ignores important errors or the prior excludes relevant possibilities.

Three practical examples

A vision system updates the probability of a candidate object class after observing image features.
A camera-pose estimator combines measurement evidence with prior information about plausible motion.
A restoration model represents several clean images that could explain the same noisy observation.

Think of it this way

Think of revising a set of working hypotheses after new evidence arrives. An explanation that predicted the evidence well gains relative weight, while one that predicted it poorly loses weight. The analogy should not be mistaken for unrestricted intuition: a Bayesian posterior follows a specified prior and likelihood, and omitted explanations cannot receive probability unless the model includes them.

Worked example

Suppose two possible surface types, A and B, have prior probabilities 0.6 and 0.4. A measured visual feature occurs with probability 0.2 under A and 0.8 under B. The unnormalized posterior weights are 0.6 × 0.2 = 0.12 and 0.4 × 0.8 = 0.32.
Their sum is 0.44. After normalization, P(A | feature) = 0.12/0.44 ≈ 0.273 and P(B | feature) ≈ 0.727. The evidence reverses the prior preference, but does not make B certain. These illustrative probabilities are assumed for the example, not measured values for a real inspection process.

Illustrative case study

In a fictional sorting project, a team uses a Bayesian model to distinguish two materials from controlled images. The model produces uncertain posterior probabilities when reflections obscure the relevant visual feature. The team routes those cases for another observation rather than forcing every item into the most likely class. It later checks whether the model's probabilities behave sensibly on representative labeled examples. The case illustrates using uncertainty in a workflow, without implying that a probability output is automatically calibrated or that the assumed material model is complete.

Common misconceptions and mistakes

A posterior is not the same as a likelihood. The likelihood describes evidence given an explanation; the posterior describes explanations given evidence, after incorporating the prior. A posterior probability of 0.9 is not a universal guarantee independent of the model. Choosing the maximum also hides the uncertainty and any competing modes. Another mistake is treating every neural-network confidence score as an explicitly derived Bayesian posterior. A score can be useful while requiring separate interpretation and calibration.

Business use

When posterior uncertainty informs decisions, document the assumptions and the consequences of incorrect decisions. Evaluate calibration, sensitivity to priors and behavior under unfamiliar data. Distinguish uncertainty about model parameters from uncertainty in a future observation. Present decision-relevant summaries rather than unexplained probability precision. If an automated workflow needs an abstention or review option, test that option against the actual costs of mistakes and additional observations.

Check your understanding

A class has a high likelihood for the observed feature. Is that alone enough to determine its posterior probability?

Answer: No. The prior and the probabilities of the evidence under competing explanations are also needed for the update and normalization.

Related terms

Prior (regularizer)
Likelihood
Evidence
Calibration
Posterior predictive distribution
Bayes' rule
prior
posterior mean
maximum a posteriori estimation
credible interval

Learn next

Learn likelihood and prior sensitivity next. Repeat a two-hypothesis update with different priors, then examine how the posterior changes when evidence is strong, weak or compatible with several explanations.

Sources

Stan, official posterior-analysis documentation: https://mc-stan.org/docs/reference-manual/analysis.html Stan, posterior distribution tools and summaries: https://mc-stan.org/posterior/ Primary documentation search excerpts consulted; full source and editorial review remain pending.

Previous: Post-training Next: Posterior (prediction)