Empty canvas mode. Pick how many classes you want, then click on the canvas to draw points. Train as normal. Tip: add ~15+ points per class for usable results.
2
300
0.10
draw points on canvas
Click the canvas to add points of the selected class. On synthetic datasets these augment the data; on blank they are the dataset.
0
02 // training
adam
0.010
16
tanh
0.00
0.000
03 // stop conditions
Any active condition will halt training when met. Status will show which one fired.
95%
5000
2000
04 // run
decision boundary live · click to add points
loss curve log-scale, last 400
05 // architecture
+ weight- weight
06 // telemetry
idle
2
0
—
—
—
—
0
07 // info
Multi-class classifier with softmax output, trained with mini-batch SGD / momentum / Adam on categorical cross-entropy. Forward + backward pass implemented from scratch. Decision boundary sampled at 64×64 grid per frame. Dropout active only at train time.