Convolutional Neural Network Simulator

Kernel / Filter

Stride 1

Scan speed 6 cells/s

Feature map

–

Kernel position

–

Current value

–

Pooled size

–

How it works

A small 3×3 kernel slides across the input pixel grid. At every position it multiplies its nine weights element-wise against the pixels underneath and sums the result into a single number — one pixel of the output feature map. Different weight patterns pick out different structure: horizontal or vertical edges, sharpened detail, or a soft blur. Stride controls how many pixels the kernel jumps between positions, shrinking the output as it grows. Max-pooling then compresses the feature map further by keeping only the strongest activation in each 2×2 block, giving the network a coarser, more robust summary that is less sensitive to small shifts in the input.

feature[i,j] = Σ(u,v) kernel[u,v]·input[i·s+u, j·s+v]
pooled[i,j]  = max( feature[2i:2i+2, 2j:2j+2] )

Filter — the fixed 3×3 weight pattern; a real CNN learns these weights from data instead of using hand-picked ones like here.
Stride — how far the kernel moves per step; stride 2 skips every other position, halving the feature-map resolution.
Max-pool — downsamples the feature map by taking the maximum value in each non-overlapping 2×2 block.
Scan speed — how fast the kernel sweeps across the grid in this visualization; the values themselves don't depend on it.

Real-world relevance: stacking many convolution + pooling stages like this — each one learned rather than fixed — is exactly how CNNs build up from simple edges to complex object recognition in image classifiers, medical scanners and self-driving perception systems.

Input pixels Kernel window Feature map Pooled map

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.