Think Together · Start with a Short Story
The robot gets an answer wrong. The girl asks, “Which dials should we adjust, and by how much?” A training algorithm uses the error to calculate a direction.
Lay Out the Process
- Forward passCalculate a prediction
- LossMeasure the difference
- BackpropagationCalculate how parameters matter
- Optimizer updateMake a small adjustment
Neural network training often uses backpropagation to calculate loss gradients for parameters, then an optimizer updates them.
You can first think of a gradient as a numerical clue about how a small parameter change may affect the loss. The model is not describing its mistakes to itself in words.
After an update, the model is checked with data again. The overall goal is to improve its performance on the task.
Now You Try
If the loss gets smaller, can you immediately say the model is more reliable in every real situation?
Ready? Open this to see the explanation
No. Loss measures a particular goal on particular data. Also check unseen examples and performance on the real task.
One Idea to Take Away
Neural network training often uses backpropagation to calculate loss gradients for parameters, then an optimizer updates them.
Try covering the diagram above and explaining each step to someone in your family. If one step is hard to explain, return to that card and look again.