An interactive book
Receptive Field.
Convolutional neural networks, one idea at a time.
Each chapter is a handful of short modules with figures you can drag, slide and step through, and a few questions at the end of each one. You start with a single 3×3 kernel on a drawing of a duck and finish with a trained network reading digits you draw.
Progress0/36 modules · 0/9 chapters
passed skipped to do
Table of Contents.
- 01Why not a dense layer?4 modules
Flatten an image and watch the parameter count explode.
- 02Convolution by hand4 modules
Slide a kernel and compute each output as a dot product.
- 03Stride, padding and dilation4 modules
Three settings that decide where the kernel lands and how big the output is.
- 04Channels4 modules
Why kernels are really three-dimensional.
- 05Pooling4 modules
Max and average pooling, and which input wins.
- 06Receptive fields4 modules
Stack convolutions and pooling layers, and watch how much of the image one output can see.
- 07Backprop through a convolution4 modules
Where the gradients go.
- 08Upsampling and transposed convolution4 modules
Spreading values back out, and checkerboard artifacts.
- 09What the network learned4 modules
A real CNN running in the browser, layer by layer.