Is inference always Bayesian updating?

No. Conditioning on evidence is a major use, but marginalization, prediction, and optimization over assignments are also inference tasks.

Does more evidence always make a belief more certain?

No. New evidence can support an alternative or reveal ambiguity. Its effect depends on the model and observed values.

Are exact methods always practical?

No. Computational difficulty depends on model structure, variable domains, and the query. Approximation may be necessary for large or complex models.

AI Glossary / Graphs and symbolic AI / Intermediate

Probabilistic inference

Definition

Probabilistic inference is the process of using a probability model to calculate or approximate answers about uncertain quantities. It can include updating beliefs after evidence, finding marginal probabilities, predicting observations, or identifying a most-probable assignment. The answer depends on the model and evidence, so correct calculation does not by itself establish that the model accurately describes reality.

What it means

A probability model specifies how quantities and events relate. Inference asks a particular question within that model. For example, it may ask how likely a hidden equipment state is after an alert, or how likely a future delay is given current workload.
Inference is distinct from learning, which estimates a model or its parameters from data, although statistical learning can itself involve inference. It is also distinct from decision-making. Knowing the probability of a delay does not alone determine whether to reroute a delivery; that decision also requires an account of available actions and their consequences.

How it works

Define query variables, observed evidence, and unobserved quantities that must be accounted for. Conditioning restricts attention to the evidence, while marginalization sums or integrates over variables not needed in the final answer. Normalization turns relevant joint weights into a conditional probability distribution.
Exact methods, such as variable elimination in suitable discrete models, exploit factorization to avoid some repeated work. Approximate methods, including sampling and variational approaches, trade computational effort against a particular kind of approximation. The requested operation matters: summing over possibilities and choosing the largest-probability possibility answer different questions and generally cannot be substituted for one another.

Three practical examples

A monitoring system can update the probability of an equipment issue after receiving a new sensor alert.
A document-processing model can infer a distribution over possible labels when text is ambiguous or incomplete.
A planning simulation can combine uncertain activity durations to estimate the probability of completing a task before a specified deadline.

Think of it this way

Think of narrowing a set of weighted possibilities after receiving a clue. Possibilities consistent with the clue are reweighted, and unanswered details are still accounted for rather than simply discarded. The analogy is limited because formal inference follows explicit probability operations; an intuitive story about a clue may overlook dependencies or the prevalence of alternative explanations.

Worked example

Consider an invented scanner model. A document has a smudge with prior probability 0.05. The scanner raises an alert with probability 0.9 when a smudge is present and with probability 0.1 when no smudge is present.
The joint probability of smudge and alert is 0.05 × 0.9 = 0.045. The joint probability of no smudge and alert is 0.95 × 0.1 = 0.095. Summing these alternatives gives P(alert) = 0.14.
After observing an alert, the smudge probability is P(smudge|alert) = 0.045/0.14, approximately 0.3214. The complementary posterior is 0.095/0.14, approximately 0.6786. The alert raises the smudge probability from 5 percent to about 32.14 percent, but no smudge remains more likely because alerts also occur in the much more common no-smudge cases.
The value 0.9 is P(alert|smudge), not P(smudge|alert). Reversing those conditions would answer a different question. The calculation is exact for the stated two-state model; whether the probabilities match a real scanner would require separate measurement and validation.

Illustrative case study

In a fictional digitization project, staff interpret an alert as a near-certain document defect because the detector often alerts on defective examples in a demonstration set. An analyst examines the full model, including how common defects are and how often clean documents trigger alerts.
The team reports posterior probabilities separately from a review decision and evaluates both on representative documents. It also checks whether the defect rate changes across incoming collections. The case illustrates why a detector's sensitivity alone cannot determine the meaning of a positive alert; no actual accuracy improvement is claimed.

Common misconceptions and mistakes

Do not equate exact inference with a true conclusion. Exactness describes calculation within a model, while model adequacy is another question. Likewise, an approximate answer can be computationally precise-looking without having a small or well-characterized approximation error.
Avoid multiplying evidence likelihoods unless the required conditional independence holds. Two copied alerts are not automatically two independent observations. Also distinguish the most-probable joint assignment from individual marginal probabilities. Finally, probability-zero evidence under a discrete model cannot be handled by ordinary conditional normalization without addressing the model or evidence specification.

Business use

Probabilistic inference can make uncertainty explicit in operational systems and support more informed review or planning. Its usefulness comes from answering a clearly defined question, not from adding a probability-looking number to an otherwise opaque output.
Record the model version, evidence used, query definition, inference method, and relevant approximation diagnostics. Test calibration and sensitivity to uncertain inputs. For important actions, connect the inferred distribution to a separate decision policy reflecting consequences and resource constraints. A high probability alone is not a complete instruction to act, and a low-probability outcome can still matter when its consequences are substantial.

Check your understanding

In the worked scanner model, what is the probability of a smudge after no alert?

Answer: P(smudge and no alert) = 0.05 × 0.1 = 0.005, while P(no alert) = 0.86. The posterior is 0.005/0.86, approximately 0.00581, or 0.581 percent.

Related terms

Learn next

Next, change the prior smudge probability to 0.20 while retaining the same alert probabilities. The alert probability becomes 0.20 × 0.9 + 0.80 × 0.1 = 0.26, and the smudge posterior becomes 0.18/0.26, approximately 0.6923. Explain why identical detector behavior gives a different posterior in a different population.

Sources

Stanford CS228 instructor-linked course notes, Variable Elimination, marginalization, factor operations, and introducing evidence: https://ermongroup.github.io/cs228-notes/inference/ve/ . Relevant sections were read. The scanner example and revised-prior exercise are original probability calculations, not measured detector performance.

Previous: Probabilistic Horn abduction Next: Probabilistic inference by automatic differentiation