Machine learning visualisations

Visualisations About

All visualisations/Deep learning

How does a CNN turn pixels into features?

Difficulty: Intermediate
Estimated time: 6 min

Loading visualisation

Step 1 of 4Read pixels locally

Inspect the highlighted 3×3 receptive field inside the 8×8 image.

A convolutional unit sees a local patch, not the complete image.

Convolutional neural network feature-map calculationUse the arrow keys to move the selected output cell and its matching three by three input patch.

vertical filter. Selected output cell row 1, column 2. Patch dot product -1. Raw response shown.

Filter

Click an output cell to reveal the exact 3×3 input patch that produced it.

Swipe diagram →