Answer key and discussion notes
Print or save as PDF: use ⌘P on Mac or Ctrl+P on Windows/Linux.
This document accompanies the student worksheet. The answers below are examples of useful explanations, not a script to memorize. Ask for reasoning, an example and a supporting source.
1. Perceptron and Deep Blue
Perceptron: in the late 1950s, Frank Rosenblatt explored a machine that learned to classify inputs by adjusting connection weights from examples.
Deep Blue: IBM’s specialized system defeated Garry Kasparov in a match in 1997. It combined large-scale search, evaluation, databases and expert input. It was neither a deep-learning model nor general intelligence.
Discuss: being excellent at one game does not mean being able to learn every task independently.
Sources: Cornell on Rosenblatt; IBM on Deep Blue.
2. Data, training and inference
- Data: example cat and dog images, with categories when using supervised learning.
- Training: adjusting parameters to improve predictions on examples. Receiving a new photograph is not, by itself, retraining.
- Inference: using the trained model to classify the new photograph.
- Possible error: blur, underrepresented animals, a different setting, dataset bias or a model limitation. Accept any well-explained example.
Discuss: more compute can help training, but it does not guarantee representative data or a correct answer.
Source and extension: how AI works, with links to original sources.
3. A quotation and an email
Quotation: ask for the reference, open the original document, and check the author, exact wording and context. A URL in the answer is not enough.
Email: review the content, identify recipients, check personal information, and obtain explicit authorization from the responsible person before sending. The agent must not expand the audience on its own.
Discuss: being able to draft a message is different from being permitted to send it.
Sources: Turing’s original paper; NIST AI Risk Management Framework.
Observe understanding
| Goal | Observable evidence |
|---|---|
| Historical milestone | The learner connects a period, a system and its mechanism without overgeneralizing its abilities. |
| Training / inference | The learner explains which process changes the model and which uses it on a new input. |
| Justified precaution | The learner proposes a source check or permission that fits the consequence. |
Use “revisit”, “explained with support” or “explained independently” to guide discussion. The game score alone is not a learning assessment. This resource and the suggested timing have not been validated in a classroom study.
Adapt the session
For beginners, focus on two or three rooms. Allow spoken answers and pair work. Use the guides or paper worksheet when spatial navigation is a barrier. These exercises do not require learners to submit personal information to an AI service.