Filter

Definition

A filter is a learned pattern used in a convolution.

It is applied to small local regions of the input.

Basic idea:

filter + local input region -> response value

A filter detects whether a local pattern is present.

---

What A Filter Learns

A filter learns weights.

Those weights define what local pattern it responds to.

Example:

filter weights:
some values positive
some values negative
some values near zero

The learned weights determine what kind of local structure produces a high response.

A filter is a trainable local pattern detector.

---

One Filter, Many Positions

The same filter is applied across many positions.

position 1 -> response
position 2 -> response
position 3 -> response

This means the filter can detect the same pattern wherever it appears.

A filter is reused across the input.

---

Multiple Filters

A convolutional layer can use many filters.

Each filter can learn a different pattern.

Example:

filter 1 -> detects pattern A
filter 2 -> detects pattern B
filter 3 -> detects pattern C

Together, they produce multiple feature maps.

Multiple filters let a layer detect multiple local patterns.

---

Final Summary

• A filter is a learned local pattern.
• It is used inside convolution.
• It produces response values at different positions.
• The same filter is reused across the input.
• Multiple filters detect multiple kinds of patterns.
• Filters are learned through training.