Loading...
Playground
Apply filters to K-culture patterns and watch feature maps change in real time!
Which K-culture pattern to try?
16×16 pixel patterns are converted to grayscale before convolution. Taegeuk and Dancheong share themes with the NN Classifier datasets.
Pick a preset or click cells to edit
The kernel slides over the input; each output cell is a weighted sum of a 3×3 neighborhood. Sobel finds edges, blur smooths, sharpen boosts outlines.
Presets
Sobel X — Horizontal brightness change — emphasizes vertical edges
Click a kernel cell to cycle values from −2 to 2.
Feature maps at each stage — play to fill conv, ReLU, and pooling cell by cell
Hover the input to highlight the 3×3 kernel window in purple.
Input
16×16
16×16 raw pixels — K-culture pattern as grayscale CNN input
Conv
14×14
3×3 filter sliding — weighted sum extracts edges, textures, spatial features
ReLU
14×14
max(0, x) — zeros out negative responses, keeps active features
Max pool
7×7
2×2 max — shrinks the map and adds slight shift invariance
Heatmap: purple = positive, orange = negative · darker = stronger