Sample glyphs
47 real images from the dataset, at the 32×32 resolution the model sees. Pick one.
Input
An 8×8 grey patch slides across the image in 64 positions, one forward pass each. Warmer areas are where hiding those pixels cost the predicted class the most confidence. Read it with some caution: a flat grey square is itself unlike anything in the training data, so even patches over blank paper disturb the prediction a little.
Prediction
Top 5 of 80 classes.
Feature maps
Every channel each convolution produces for this glyph. Each tile is normalised independently, so bright means “strong response relative to this channel’s own range”. A completely black tile is a channel that stayed silent on this input — after ReLU, most of these maps are exactly zero.
Draw a glyph yourself optional
The selected gallery glyph is ghosted behind the pad as a tracing guide. Predictions update as you draw.
Worth being honest about this: the model was trained on isolated, evenly scanned dataset glyphs. Freehand mouse or finger input has different stroke weight, centring and noise, so it is out of distribution. Expect predictions here to be noticeably worse than the 90.2% test accuracy measured on held‑out dataset images.