A picture is a grid

Begin with the mug

In the opening picture, the mug feels like one object. In an image file it is spread across many samples. Each pixel records numerical values, often red, green and blue channels. Those values are influenced by illumination, material, camera settings and processing. The machine does not receive the word “mug” along with those pixels. A learning system has to connect patterns in the numbers with the task we ask it to perform.

Look at a small neighbourhood

Before thinking about a large neural network, take three neighbouring brightness values. A simple filter can compare the left and right values. If they differ sharply, you may be near an edge. Slide the same filter across a row and you obtain a sequence of responses. This is the basic intuition behind a local operation: each output looks at a small region, rather than trying to interpret the entire scene at once.

Edges are clues, not objects

An edge response could come from the mug boundary, a shadow, the notebook or an exposure change. It does not tell you which one. That is why a single filter is useful but insufficient. In a learned vision system, later operations combine local responses into representations that can support a task. Even then, a feature that helps on training images may fail when the background or lighting changes. A sharp response is evidence of contrast, not understanding.

Do one calculation slowly

Use the row [0, 0, 1, 1, 1] and the three-value filter [−1, 0, 1]. Multiply matching entries and add, then shift one place. We will use cross-correlation, which does not reverse the filter; many deep-learning libraries call this operation convolution. Keeping that convention explicit prevents a sign surprise. Once this tiny example feels ordinary, a grid of filters and multiple image channels is a manageable extension rather than a leap of faith.

SEE THE RELATIONSHIPS

From structure to behavior.

MAKE THE IDEA PRECISE

The mathematics.

Build the equation, step by step

1. Align the filter with three consecutive samples. Multiply each sample by its corresponding weight.
2. Add the products to obtain one output. For [0,0,1], the result is 1.
3. Move the window by one sample. With length 5, filter width 3, stride 1 and no padding, there are 3 valid outputs.

WORK THROUGH THE NUMBERS

Worked example

The three windows produce [1, 1, 0]. The first two see a change from dark to bright; the last window is constant and gives zero. Reversing the filter reverses the sign of the edge response.

A small Python experiment.

Python 3 · Standard library only · Synthetic educational example. Run this in your own notebook or a local file.

row = [0, 0, 1, 1, 1]
kernel = [-1, 0, 1]
result = [sum(row[i+j]*kernel[j] for j in range(3)) for i in range(3)]
print(result)

Expected output

[1, 1, 0]

Change one quantity, predict the result, and then run the code again. Explain any difference before adding complexity.

Read with the original sources.

These readings support the concepts and deeper investigation. The worked numbers and illustrations above are teaching examples.