What does a feature map look like early versus late in a network?

Early maps respond to edges and colors; deeper maps respond to parts and whole objects.

Is a heatmap explanation a feature map?

Closely related: saliency heatmaps are derived from feature maps and gradients to show what drove a decision.

Do feature maps exist in text models?

Analogs exist in attention patterns, though the sliding-window map structure is native to images and signals.

AI Glossary / AI and machine learning / Intermediate

Feature map

Definition

A feature map is the output produced when a learned detector is applied across an input, such as the grid of responses an edge detector leaves at every position of an image.

What it means

Inside a vision network, each layer holds many detectors, each looking for one pattern, an edge, a texture, a curve. Sliding one detector across the whole image produces a map: high response where the pattern appears, low where it does not. That grid of responses is a feature map.

Stack the maps from many detectors and you have the layer's full reading of the image: not pixels any more, but pattern-presence recorded at every location.

The same idea appears beyond images. Any transformation that re-expresses data in a new space, including the mathematical maps behind kernel methods, is called a feature map: the old input seen through a new lens.

Reading feature maps is one of the few ways to see what a network notices. Visualizing them shows which patterns each layer hunts, and early layers' maps famously look like edge and color detectors.

For a business, feature maps matter mainly as an explanation tool: when an image model must be defended, showing what it detected, and where, is more persuasive than quoting its accuracy.

How it works

Apply a detector, a small learned filter, at every position of the input, producing one response per position.

Collect the responses into a grid aligned with the input: that grid is the feature map for that detector.

Repeat with many detectors in parallel, so the layer outputs a stack of maps, one per pattern it has learned to hunt.

Pass the stack to the next layer, whose detectors hunt patterns of patterns, building parts from textures and objects from parts.

For inspection, render selected maps as images: bright regions show where in the input that detector fired.

When visualizing maps for a business audience, annotate what each bright region corresponds to in the original image; an unannotated map is modern art, an annotated one is evidence.

Three practical examples

Quality control: a defect-detection model's maps highlight the regions where scratch-like patterns fired.
Medical imaging: maps show which areas drove the model's suspicion, supporting, not replacing, the radiologist.
Document analysis: layout maps locate tables, logos and signature blocks before content is read.

Think of it this way

Think of a search party walking a field, each member hunting one thing, footprints, broken branches, discarded litter, and radioing in where they find it. Each member's report over the field is a feature map: one pattern, everywhere it occurs.

The limit: the search finds only what the searchers were told to hunt. Feature maps record the presence of learned patterns, and a pattern the network never learned leaves no map at all.

Worked example

explaining a defect flag

Suppose a factory's vision model flags a machined part as defective, and the line manager wants to know why before stopping production. The figures are illustrative.

The team renders the feature maps of the layer that fired strongest and finds responses concentrated on the part's inner rim.

Closer maps reveal the trigger pattern: a fine circular scratch texture the model learned to associate with a worn tool.

The tool is replaced, the flags stop, and the explanation, not just the alert, is what the manager remembers.

Illustrative case study

A fictional retailer bought an image moderation system and discovered, only after a dispute, that its strongest feature maps fired on image backgrounds rather than products, because the training photos of rejected items had shared a distinctive backdrop.

The model had learned the backdrop, not the violation.

The invented lesson: a model's maps show what it actually looks at. That can differ embarrassingly from what the brochure says it looks at.

Common misconceptions and mistakes

A feature map is a visualization for humans. It is an internal data structure the network computes with; visualizing it is optional, its existence is not.
Feature maps only exist for images. The sliding-detector structure is natural for images and audio, and the broader idea, data mapped into a new space, appears throughout machine learning.
Bright regions prove the model is right. They show what the model attended to, which is evidence about its reasoning, not proof of its correctness.

Business use

Ask vendors of image systems for feature-map or attention evidence on your own samples, not just accuracy numbers.

Use map inspections when debugging: a model looking at the wrong region is diagnosed faster visually than statistically.

Keep the caveat: maps show attention, not justification, and a confident map over the wrong region is still a wrong model.

Use map evidence in supplier acceptance tests: a defect detector should fire on defects in your photos, and the maps show whether it does before you trust its alerts.

Check your understanding

What does a feature map record about one detector in a vision network?

Answer: It records, at every position of the input, how strongly that detector's pattern was present, turning pixels into a grid of pattern-presence.

Related terms

Learn next

AI for Business Owners explains how to ask for evidence of what an image model actually looks at: https://aiforbusinessandlife.com/academy/ai-for-business-owners. [C2]

Sources

[S1] Russell and Norvig, AIMA glossary; [S12] Aalto University, Dictionary of Machine Learning. Full URLs are in the source ledger. Verified 11 October 2026.

Previous: Feature learning Next: Feature matrix