Kernel

Definition

A kernel is the small set of weights used by a convolution.

It defines the size and values of the local pattern being applied.

Example:

3 values in a row
3 by 3 values in a grid

A kernel is the weight structure of a convolution filter.

---

Kernel vs Filter

The words kernel and filter are often used closely.

For this chapter:

kernel:
the small weight window

filter:
the learned pattern detector using that kernel

In many practical discussions, people use them almost interchangeably.

Kernel emphasizes the weight window; filter emphasizes the pattern detector.

---

Kernel Size

Kernel size determines how much local context is used at once.

Examples:

small kernel:
uses a small neighborhood

larger kernel:
uses a larger neighborhood

A larger kernel can see more local information, but it also uses more parameters.

Kernel size controls local context.

---

Sliding The Kernel

The kernel is moved across the input.

At each position:

kernel weights combine with local input values
produce one response

Repeating this across positions creates a feature map.

A convolution applies the kernel repeatedly across the input.

---

Final Summary

• A kernel is a small weight window used in convolution.
• It defines the local pattern size.
• Kernel and filter are closely related terms.
• Kernel size controls how much local context is used.
• The kernel slides across the input.
• Its responses form a feature map.