A tiny convolutional network classifies 28×28 blood-cell images in your browser. Click any feature map to watch convolution, ReLU, pooling, or softmax actually happen. Inspired by Polo Club’s CNN Explainer, with a hematology example instead of everyday objects.
These are real BloodMNIST test images (white-blood-cell / platelet tiles). The model never sees your own scans — there is no upload, on purpose.
Hover a square to see which previous maps feed it. Click a square to open the math. Red–blue tiles are activations. The yellow–green strip under each conv filter is that filter’s 3×3 kernels — one kernel per incoming channel. Max-pool uses a 2×2 window of 4 activations (no learned weights) and keeps the largest. ReLU and max-pool do not mix channels; only convolution sums across them.
A 3×3 kernel looks for the same edge, granule, or halo wherever it appears. That is why CNNs scale to images far better than dense nets. See parameter growth.
Max-pool slides a 2×2 window (four numbers, no learned weights) and keeps the largest. That is why a lymphocyte still counts if it is not perfectly centered — and why the map gets half as wide.
The bars sum to 1, but they are not calibrated risks. A 0.91 “lymphocyte” score is not a 91% chance. See calibration & thresholds.
Bright pixels mean “this filter fired,” not “this is a nucleus” unless you verify it. For how explanations go wrong in health ML, see interpretability.