AI Glossary / Computer vision and multimodal AI / Intermediate

Preimage

Definition

A preimage is an input, or a set of inputs, that maps to a specified output under a function. In machine learning and vision, a preimage problem often asks which original-space input could produce a given feature representation.

What it means

A forward mapping turns an input into an output: an image becomes a feature vector, for example. Looking for a preimage reverses the question. There may be exactly one compatible input, many compatible inputs or none. The mathematical preimage of an output set includes every input mapped into that set and is defined even when an ordinary inverse function does not exist. In feature-based vision, the term often refers to reconstructing an input that matches or approximately matches a target representation. This should not be confused with simply retrieving the original image from storage.

How it works

Given a mapping f and a target y, seek x such that f(x) = y. If exact equality is impossible or difficult, an algorithm may minimize a feature-space discrepancy, often with constraints or a prior encouraging plausible inputs. The result is then an approximate preimage under a specified criterion.
Information-losing mappings make the problem ambiguous. For example, averaging pixels discards their individual arrangement. A target created by manipulating a feature vector may also lie outside the set of representations produced by any valid input. Optimization can still return a nearby candidate, but it should not be described as an exact inverse without evidence. Initialization and regularization can influence which candidate is selected.

Three practical examples

• A feature-visualization experiment searches for an image that produces a chosen pattern of neural-network activations.
• A kernel-method study seeks an approximate input-space reconstruction from a feature-space representation.
• An image-processing lesson lists all small input arrays that share the same average value, demonstrating information loss.

Think of it this way

Think of receiving a total price and asking which items produced it. Several baskets may have the same total, and some totals may be impossible under the available prices. The analogy explains ambiguity and nonexistence. It does not mean that every image representation is a simple sum or that a mathematically compatible reconstruction is the actual original.

Worked example

Let f(x) = x² over the real numbers. The preimage of the output 4 is the set {−2, 2}. The preimage of 0 is {0}, and the preimage of −1 is empty. These outcomes show why a preimage need not be a single value.
For an imaging-style example, let f(a,b) = (a+b)/2. The two-pixel inputs (0,100), (50,50) and (100,0) all map to 50. Knowing only that output cannot reveal which input occurred. A smoothness preference might select (50,50), but that choice comes from an added assumption rather than information contained in the average.

Illustrative case study

In a fictional visual-search project, a team asks a reconstruction tool to generate an image matching a stored feature vector. Several runs produce different images with similar feature-space scores. The team presents them as model-based illustrations of the representation, not recovered originals. It also checks whether apparent matches remain meaningful under the application's actual retrieval evaluation. This case demonstrates how preimage ambiguity affects interpretation without implying that the representation necessarily reveals or uniquely identifies a source image.

Common misconceptions and mistakes

The notation f⁻¹ for a preimage set does not automatically imply that f has a single-valued inverse. An approximate feature match is not proof of exact equality. A plausible-looking candidate is not proof that it was the original input, and a low feature distance does not necessarily preserve every semantic property. Another mistake is assuming that every edited feature vector corresponds to some valid image. The representation may occupy only a restricted subset of its surrounding numerical space.

Business use

When using reconstruction or feature visualization in a product, describe what is being recovered and which assumptions select the result. Distinguish an exact inverse, an approximate optimization result and a generative illustration. Evaluate agreement in both feature space and any relevant input-space or task criteria. Preserve information about the mapping and reconstruction settings so users do not mistake a model-dependent candidate for an observed record. The usefulness of a preimage depends on the purpose of the representation.

People also ask

Can a preimage contain more than one input?

Yes. Any many-to-one mapping can send several inputs to the same output.

Can a target have no preimage?

Yes. A target outside the mapping's attainable outputs has no exact preimage in the stated domain.

Does regularization remove ambiguity?

It can select or favor a candidate, but it adds assumptions rather than restoring information that the mapping discarded.

Check your understanding

Two different images produce the same feature vector. A reconstruction method returns one of them. Has it proved which image was originally observed?

Answer: No. The shared feature vector does not distinguish the two inputs, so additional evidence would be needed.

Related terms

• Feature map
• Inverse problem
• Regularization
• Function
• inverse image
• inverse function
• kernel PCA
• representation inversion
• non-identifiability

Learn next

Learn many-to-one mappings and inverse problems next. Start with a simple averaging function, identify several compatible inputs, and then examine how adding a smoothness preference changes the selected reconstruction without changing the evidence.

More in Computer vision and multimodal AI

• Power spectrum - A power spectrum describes how a signal's power or squared-amplitude content is distributed across frequencies.
• Preconditioning - Preconditioning transforms a numerical problem into an equivalent or suitably related form that is easier for an iterative solver to handle.
• Prefilter (anti-aliasing) - An anti-aliasing prefilter reduces signal content that a subsequent sampling step cannot represent reliably.
• Prim's algorithm - Prim's algorithm finds a minimum spanning tree of a connected, weighted, undirected graph.
• Prior (regularizer) - A prior is a model's probability distribution or preference over possible solutions before incorporating the current observation.
• Probabilistic relaxation labeling - Probabilistic relaxation labeling is an iterative method for refining uncertain labels using compatibility relationships among neighboring items.

Sources

Cornell University, function notes defining preimages of sets: https://www.cs.cornell.edu/courses/cs2800/2015sp/handouts/jonpak_function_notes.pdf Bishop, Pattern Recognition and Machine Learning, discussion of preimages in kernel methods: https://www.microsoft.com/en-us/research/wp-content/uploads/2006/01/Bishop-Pattern-Recognition-and-Machine-Learning-2006.pdf Primary teaching and author-textbook search excerpts consulted; full source and editorial review remain pending.

Previous: Prefrontal cortex basal ganglia working memory Next: Premise