Start with random weights. Use many labeled examples so the model can improve.
A few of its flashcards
Hidden neurons — model size32
Learning rate — how bold each tweak is0.003
Training rounds — passes over the data4
Test accuracy
—
Loss
—
Round
0/4
Accuracy ↑
Loss ↓
2 · Test it — draw a digit
Draw any number from 0 to 9. The model turns your drawing into a 28×28 image and predicts a label.
what the model receives
Hidden neurons firing
?
Draw a digit to see the prediction.
Try this: Does a larger model always do better? What changes if you train for only 1 round?
Draw a messy 7 or a 1 with a slash. Which digits are easy to confuse?