How does it differ from open-set recognition?

Open-set recognition addresses unknown classes at inference; open world approaches also address incorporating new classes over time.

Does it name an unknown class immediately?

Not necessarily. Semantic labels usually require additional information or supervision.

Is it the same as zero-shot classification?

No. Zero-shot methods may classify specified unseen categories using descriptions or other information, while open world settings include categories not yet specified.

AI Glossary / Probabilistic machine learning / Advanced

Open world classification

Definition

Open world classification recognizes examples from known categories, allows unfamiliar examples to remain unknown, and supports adding newly established categories over time. It combines classification with mechanisms for handling novelty and updating the class set. It does not assume that every future input belongs to a category available during initial training.

What it means

A conventional closed-set classifier must choose among its existing labels. An open world system needs another possible outcome: the available evidence does not justify assigning a known class. That outcome can lead to review, collection of additional examples and eventual introduction of a new category.
Unknown is not itself a complete semantic explanation. It can reflect a genuinely new category, a damaged measurement, an unusual example of a known class or a limitation of the representation. Establishing and naming a new class requires more evidence than merely rejecting one input.

How it works

The system first represents inputs and evaluates their compatibility with known classes. A rejection mechanism determines whether a known-class decision is sufficiently supported. This might involve distances, calibrated scores or a specialized novelty model; the appropriate design depends on the application.
Rejected examples are examined and, when justified, organized into new labeled categories. The classifier is updated to recognize those categories while continuing to handle future unknowns. The update must also be checked for deterioration on previously learned classes.
Evaluation should therefore include known-class recognition, unknown rejection and performance after class additions. A method that rejects every input can achieve perfect rejection of unknowns while being useless on known classes. Conversely, excellent closed-set accuracy can conceal forced labeling of entirely unfamiliar inputs.

Three practical examples

Product catalogs: a visual classifier recognizes stocked product types and flags unsupported types. Reviewed examples can later support adding a new product class.
Document routing: a service assigns familiar request categories while referring unfamiliar patterns for examination. A new route is introduced only after a meaningful category is established.
Robotics research: an object-recognition system distinguishes familiar categories from unsupported observations and evaluates how adding classes affects old recognition results. Rejection alone does not provide safe action planning.

Think of it this way

Imagine a librarian with an expandable catalog. Familiar books can be filed under existing headings; unfamiliar material can wait for review. Creating a useful new heading requires understanding the material and checking how it relates to existing headings. Simply putting every difficult book in a box marked unknown would not complete the cataloging process.

Worked example

Consider an original one-dimensional teaching system. Known class A has prototype 0 and known class B has prototype 4. Accept the nearest prototype only when its absolute distance is at most 1; otherwise return unknown. This is a simplified teaching rule.
An input at 0.6 is assigned A because its nearest distance is 0.6. An input at 3.4 is assigned B for the same reason. An input at 7 is rejected because its distances from A and B are 7 and 3, both above the threshold.
Suppose reviewed examples justify a new class C with prototype 7. Adding that prototype lets the system assign an input at 7.4 to C because the distance is 0.4. An input at 10 still remains unknown: its nearest prototype is 3 units away. The class addition expands what is known without closing the world to further novelty.
Now consider a stipulated evaluation batch before adding C. It contains eight known-class examples and four genuinely unknown examples. Seven known examples receive correct accepted labels, while one is rejected. Of the four unknown examples, three are rejected and one is incorrectly accepted as a known class.
Known-class correct acceptance is 7/8, or 87.5%. Unknown rejection is 3/4, or 75%. The unknown false-acceptance rate is 1/4, or 25%. Among the four rejected examples, three are truly unknown, so the unknown fraction in the review queue is 75%. Treating correct known labels and correct unknown rejections as successes gives 10/12, approximately 83.3%, but that aggregate hides the different error types. These counts are stipulated illustrations, not results from the three individual points above or a trained classifier.

Illustrative case study

Fictional company Fern Catalog uses image classification to organize incoming inventory. Its team reserves a review path for items not supported by the current category set. Reviewers separate new product types from poor photographs and atypical examples of existing types.
When enough labeled evidence supports a new category, the team updates the classifier and evaluates both old and new categories. It also tests additional held-out categories that remain unknown. The team tracks review workload so a stricter rejection rule does not silently overwhelm staff. This fictional workflow reports no deployment results.

Common misconceptions and mistakes

Open world classification is not simply closed-set classification with more labels. Nor is it complete once an unknown threshold has been added. Learning newly established classes is a separate part of the evolving system.
A high maximum softmax score does not guarantee that an input belongs to a known class. A rejected example is not automatically a new category, and an outlier is not necessarily a useful category seed. Finally, learning new classes without testing old ones can hide forgetting. Evaluation that exposes all future classes during tuning can also weaken claims about genuinely unfamiliar categories.

Business use

The approach is relevant when categories evolve and forcing every input into a fixed taxonomy creates avoidable errors. Its operational value depends on more than the classifier: review capacity, labeling standards, update procedures and evaluation all matter. Teams should measure both false acceptance of unfamiliar inputs and unnecessary rejection of familiar ones.

Check your understanding

Why did the input at 7 initially remain unknown?

Answer: Neither existing prototype was within the acceptance radius.

Why did adding C not make the input at 10 acceptable?

Answer: It still lay outside every class's allowed region.

Why report separate known and unknown metrics?

Answer: A combined score can conceal very different rejection and misclassification behavior.

Related terms

Open set recognition
Novelty detection
Class incremental learning
Zero-shot learning
Out-of-distribution detection
Human-in-the-loop learning
Catastrophic forgetting

Learn next

Compare closed-set predictions with explicit rejection, then evaluate a sequence of class additions. Keep some categories entirely unseen until evaluation and measure review workload alongside recognition performance.

Sources

Bendale and Boult, Towards Open World Recognition (2015), Definition 1 independently consulted: https://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Bendale_Towards_Open_World_2015_CVPR_paper.pdf . The prototype and evaluation examples and fictional case are original illustrations.

Previous: Open-weight Next: Operation list