LearnDeep learning
Convolutional network
Slide a small filter across a picture and watch it find edges: the trick that lets networks see.
Click or drag on the left grid to draw. Point at the feature map to see how one output is computed.
How it works
Small filters, shared everywhere.
A convolutional layer slides a small grid of weights, the filter or kernel, over the image. At each position it multiplies the nine pixels under it by the nine weights and adds them up, giving one number of the feature map. Point at any output to see that sum.
The same nine weights are used at every position, so the filter finds its pattern anywhere in the picture with very few parameters. An edge filter gives strong positive and negative values where brightness changes; ReLU then keeps only the positive responses, and 2 × 2 max pooling keeps the strongest value of each block, halving the size.
Here you choose classic hand-made filters. In a trained network the filters are learned by backpropagation, and deeper layers combine edges into textures, parts and whole objects.