به انگلیسی

How learned weights work with code

Computing · Artificial Intelligence · ages 14-15

Name ______________________ Date ____________

What does a weight do in the toy spam model?
- Stores a complete written explanation for every email
- Multiplies an input before the total is added
- Deletes the threshold comparison from the calculation
What distinguishes fitted weights from a hand-written spam rule?
- Weights require no code to calculate a prediction
- Weights are learned numerical parameters
- Weights are complete rules typed separately for each email
Which part stays fixed while this toy model's weights and bias are fitted?
- The current value of every fitted weight
- The current numerical value of the bias
- The multiply-and-add calculation
How could you investigate why the small toy model marked a message as spam?
- Assume fitted numbers make every calculation impossible to inspect
- Trace the inputs, weights and calculation
- Read a separate hand-written rule stored for that exact message
During training, the toy model's prediction differs from a correct example label. Which action fits its learning process?
- Remove every example whose prediction differs from its label
- Replace the whole calculation with a new one after each error
- Adjust weights and bias using the error
Training has stopped. The model now only predicts, with no parameter-update process running. What happens to its weights?
- They keep changing automatically with each new prediction
- They reset to zero whenever a new message arrives
- They stay fixed during prediction
Retraining the toy model on more examples can change its predictions without changing its fixed calculation.
Circle one: True False
A vendor says its model has learned weights, so there is no code or calculation to inspect. Which response is best supported?
- Ask how learned weights and code produce predictions
- Assume every prediction has a separate rule typed by a person
- Assume fitted weights mean there is no software to examine

Answer key

For grown-ups. Fold this page away before handing over the rest.

How learned weights work with code W1-mt_3_1toZUxak-s1

Multiplies an input before the total is added · Each input is multiplied by its weight. Those products are then added with the bias.
Weights are learned numerical parameters · Training fits numerical weights from examples. Software still defines how those weights produce a prediction.
The multiply-and-add calculation · This toy model keeps its calculation fixed while training adjusts parameter values.
Trace the inputs, weights and calculation · The toy model's inputs and parameters can be followed through its calculation to its prediction.
Adjust weights and bias using the error · Training uses the prediction error to adjust parameters. The adjustment aims to reduce errors; it is not a promise that every prediction improves.
They stay fixed during prediction · Prediction alone uses the fitted weights. The question rules out training or other parameter updates.
True · Retraining can fit different parameter values. The same calculation can then produce different predictions.
Ask how learned weights and code produce predictions · Learned weights are used by a defined calculation. Fitted parameters do not remove code that can be examined.