Follow the Learning Step

If a parameter's cost gradient is positive, which way should it move?

9 weights + 4 biases · ← negative · ↑ zero · positive →
Training ExampleSleep hStudy hKnownPredictedDifference
Try a New Input—

Inference pauses training and keeps parameters fixed · Three examples do not establish accuracy on new students

Cost: Scaled MSE

—

Targets = score ÷ 100 · Mean of 3 squared differences