Draw digits. Inspect live CNN kernels, learned weights and exact calculations.
Read full description
CNN Lab makes a real convolutional neural network visible, editable by your input, and inspectable one operation at a time.
DRAW. FOLLOW THE KERNEL. VERIFY THE MATH.
Write a digit with your finger and watch an actual on-device CNN respond. Rotate the 3D network, choose a feature-map channel, and inspect the precise input window, learned filter weights, elementwise products, bias and ReLU behind an output pixel.
GO BEYOND A FINISHED PREDICTION
Follow max-pooling regions and their selected maximum. Explore flattening, connected-neuron accumulation and softmax probabilities. Pause, move to any operation, or adjust playback from 0.5x to 128x. Displayed arithmetic is checked against actual network activations.
KEEP YOUR EXPERIMENTS
Save your input together with its layer, filter and operation position. Reopen a note to replay that exact experiment. Daily goals, learning XP and cumulative milestones give you a reason to return and explore another layer.
A PRETRAINED MODEL, ALREADY ON YOUR DEVICE
The bundled MNIST Mini CNN was trained on 60,000 samples and measured at 97.64% accuracy on a separate 10,000-sample MNIST test set. Your own handwriting may produce different results. Core inference does not require a remote AI service.
YOUR CHOICES
The core simulator is available without registration or watching ads. Optional gems come from completed rewarded ads or Google Play purchases and unlock notebook capacity and a visual theme. Purchased and ad-earned gems are recorded separately. There are no subscriptions. Consumable purchases already consumed may not be restored after app data deletion. Unfinished purchases are checked and delivered after verification.
Usage analytics is off by default and can be enabled in Settings. Drawings and calculation values stay on your device. Automatic ad formats are separately optional and limited. Korean, English, Japanese and Simplified Chinese are supported.
Built for learners, educators and developers who want to understand what a CNN actually computes.
Support: https://ikhpark.com/blog/app-support
Privacy: https://devkhpark.tistory.com/470
Terms: https://ikhpark.com/blog/with-theapp-terms-of-use-2026-08-20