A CNN Reads a Digit

A real convolutional network, 26,698 learned numbers trained on MNIST, read layer by layer: every feature map, every pooled stack, the dense layers and the ten probabilities. Then draw your own digit.

Builds onPooling and Feature Maps

Digit CNN trained on MNIST (99.0% test accuracy) · MNIST test digits, CC BY-SA 3.0

Loading the digit CNN and its test digits (98 KB)…

conv1	[8, 28, 28]	80
pool1	[8, 14, 14]	0
conv2	[16, 14, 14]	1,168
pool2	[16, 7, 7]	0
fc1	[32]	25,120
fc2	[10]	330
total		26,698

conv1	[8, 28, 28]	80
pool1	[8, 14, 14]	0
conv2	[16, 14, 14]	1,168
pool2	[16, 7, 7]	0
fc1	[32]	25,120
fc2	[10]	330
total		26,698

DSAverse

A CNN Reads a Digit

See also

A CNN Reads a Digit

See also