How CNNs See the World

A neural network can't "see" photos like you do. It reads a grid of numbers — each pixel is an intensity from 0 (black) to 255 (white). Drag the slider to see how resolution affects detail.

A 3x3 kernel slides across the image one pixel at a time. At each position, it multiplies and sums — a dot product. Drag the slider to move the kernel manually, or press Play to animate.

Each kernel is a feature detector. Click the presets above to see how different kernels produce different feature maps. ReLU zeros out negative values — if the filter didn't match, the neuron stays silent.

Max pooling slides a 2x2 window and keeps only the maximum. Try changing the stride: stride 1 overlaps windows, stride 2 halves the map. This makes the network invariant to small shifts.

A CNN stacks conv + ReLU + pool blocks. Drag the slider to add layers: early layers detect simple edges, deeper layers combine those into textures, parts, and whole objects.

Toggle between inputs above to see how predictions change. The final feature maps get flattened, fed through dense layers, then softmax turns scores into probabilities. The highest is the network's answer: Cat (91%).