CNNs Explained From the Inside: How Pixels Become Predictions

Figure 1. From raw pixels to predictions — showing how convolution, activation, pooling, deep feature extraction, backpropagation, and explainability work together inside a CNN.

When we look at an image, we immediately see objects.

A computer sees numbers.

A CNN bridges that gap.

But if we only learn the equations for convolution, ReLU, pooling, and backpropagation, we miss the most interesting question:

What do those changing numbers actually mean inside the network?

A CNN is not simply transforming numbers.

It is gradually transforming their meaning.

A pixel initially represents brightness.

After convolution, a number represents evidence for a pattern.

After deeper layers, numbers represent increasingly complex visual features.

Finally, those features become evidence for a prediction.

And through backpropagation, the network learns which transformations are useful.

So let’s follow an image through a CNN and ask at every stage:

What changed?

Why did it change?

How does that help the model learn?

Can we explain what the model is doing?

1. Start With an Image