I Built a CNN You Can Watch Think — in 3D
Most explanations of a convolutional neural network are diagrams and equations. I wanted the opposite — something you can watch. So I trained a small CNN to recognize handwritten digits, then rendered it in 3D so you can see every step as an image flows through it.
Filters: tiny pattern detectors
A CNN learns small 3×3 filters — little grids of numbers, each one a pattern detector. One fires on horizontal edges, another on diagonals, another on curves.
Convolution: slide, multiply, sum
Each filter slides across the image. At every position it multiplies its weights by the patch underneath, adds them up, and keeps the result if it's positive (ReLU). Slide across the whole image and you get a feature map — bright wherever the filter found its pattern.
Pooling: shrink, keep the strongest
Pooling then takes each little 2×2 window and keeps only the largest value, halving the size while preserving the strongest signals. That makes the network faster and more tolerant of small shifts.
Edges → shapes → an answer
Stack these steps — convolve, pool, repeat — and the network builds from simple edges up to whole shapes. A final layer turns all those features into ten scores, one per digit, and the brightest one is the answer.
Why I made it
The math tells you what to compute; it doesn't show you the thing happening. Watching a digit light up filter by filter, then collapse into a confident prediction, is what made convolution finally click for me.
Comments