A neural network that looks at an image one pixel at a time is doomed. A 224x224 colour photograph carries over 150,000 input values, and wiring every one to every neuron in a hidden layer produces billions of parameters before the model learns anything. Convolutional neural networks solve this by refusing to treat pixels as independent facts. They assume something about the world: that a pattern worth recognising in the top left corner is still worth recognising in the bottom right. That assumption becomes an architecture. - A kernel is a small grid of weights, typically 3x3 or 5x5, that slides across the image - At each position it multiplies overlapping pixel values by its weights and sums them into a single number - The result is a feature map, a new image showing where that pattern occurred - The same kernel is reused everywhere, so one filter costs nine weights rather than 150,000 This is weight sharing, and it is the whole trick. Translation invariance is built into the struc...