Models / Simple CNN / Workspace
code

Simple CNN · Live Inference

Real forward + backward passes on CPU: predictions, Grad-CAM.

Simple CNN · CPU
analytics99.2% AccMNIST60K paramsIn this workspace:Grad-CAM
RecognizesHandwritten digits, 0 through 910 classes · the digits 0–9
Accuracy99.2% Acc99.2% on the MNIST test set
Works best onOne clearly drawn digitDark strokes on a light background; converted to 28×28 grayscale.
warning
Expect confident mistakes outside its scope

This network has only ever seen handwritten digits. Give it a photo, a letter, or a scene and it will still answer with a digit, usually a confident one, because a closed-set classifier has no "none of these" option. Mistakes on anything that is not a single 0–9 digit are expected, not a defect.

Concepts in play
filter_center_focusConvolutionOne small grid of weights, slid across every position, turns pixels into feature maps by dot product.

Top-5 Predictions

Upload an image and run analysis to see predictions.

Input

image

Activation Map

LOW→HIGH
blur_on

Layer Source

—

Focus Metric

—

Latency

—

Confidence (top-1)

—

Run analysis to compute

Inference Latency

—

CPU forward pass