คอนโวลูชัน: เคอร์เนลเลื่อนไปบนพิกเซล

To a network, a photo is a grid of pixel values. A small grayscale image might be 28x28 numbers, each from 0 (black) to 255 (white).

Flattening an image into one long vector throws away the spatial layout. Pixels near each other matter, and a plain dense layer ignores that.

A kernel is a tiny grid of weights, often 3x3. It slides across the image, looking at one small patch at a time.

At each position the kernel covers a patch, multiplies overlapping values, and sums them into one number. Then it slides over and repeats. 🔍

Each weighted sum becomes a single pixel in the result. Slide across the whole image and you build a fresh grid called a feature map.

A well-tuned kernel lights up where its pattern appears. One kernel might fire on vertical edges, another on bright corners or smooth textures.

This classic 3x3 kernel responds strongly to vertical edges, where light meets dark across a column.

You do not hand-pick kernel numbers. Training adjusts them so each filter learns whatever pattern helps the network classify images.

One kernel reuses the same weights across the whole image. This weight sharing means far fewer parameters than a dense layer would need.

nn.Conv2d wraps all of this. Here a Conv2d takes 1 input channel and learns 8 different 3x3 kernels.

That layer learns 8 kernels, so it outputs 8 feature maps. Each one highlights a different pattern the network found useful.

Let us check what a convolution kernel actually produces.

You learned that a kernel slides over pixels, summing weighted patches into a feature map. Its weights are learned, shared, and tuned to spot edges and textures. 🎉