This sandbox was generated by Claude (Anthropic) for the ML 101 course materials. It applies a fixed, named convolution kernel to a small grayscale image and shows the four things the Module 7 lesson describes in words: the feature map a convolution produces, the effect of ReLU, the down-sampling of max-pooling, and how the spatial dimensions shrink at each stage. Nothing here is sketched — every feature map is drawn from the computed array, cell by cell.

These kernels are fixed, not learned. A real CNN learns its kernels by gradient descent over training data; this page does not train anything. It uses the standard textbook operators — Sobel (edges), the Laplacian, a box blur, a sharpen kernel and the identity — so you can see exactly what a convolution does, which is the prerequisite for understanding what a CNN learns. Nothing here is presented as a learned filter.

The convolution of deep learning is cross-correlation. It does not flip the kernel — the flip only changes the result for asymmetric kernels (the Sobel pair) and is a naming convention. This page follows the CNN convention (PyTorch, TensorFlow), and the equation block states it.

Computed: the convolution of the chosen kernel over the chosen image, the ReLU, the 2×2 max-pool, every output dimension, the response energy of each kernel, and the optional second convolution layer — all real arithmetic. Published figures: none — the kernels are standard operators and everything else is a direct computation. Chosen rather than computed: the synthetic grayscale patterns, the image size, and the seed for the optional pixel noise.