Stretch tasks What AI actually is

Round 1: Quiz Blitz

1. Machine learning finds patterns in examples instead of following rules a programmer wrote In machine learning the rules are not written by hand: the system finds them in the training data.
2. Collect images labelled cat or not cat → Train: adjust internal values to get more labels right → Test it on images it has never seen → Use it to classify new images Collect, train, test on unseen data, then use. Skipping the test step hides whether it has really learned.
3. False It has found statistical patterns that go with the label cat. That is why it can be confidently and bizarrely wrong.
4. To check it learned general patterns rather than memorising its training images A system that only works on the pictures it was trained on has memorised them, not learned.
5. Spam filters; Face unlock on a phone; Film recommendations on a streaming service Each of the three finds patterns in data. A switch and a simple sum just follow fixed rules.
6. The labelled examples a system learns patterns from Training data is everything the system learns from, so its quality decides how good the system is.
7. It learned that grass goes with dogs, not what a dog is It latched on to something that usually appears with dogs. Correlation is not understanding.
8. True A pattern that usually works can fail badly on something slightly unusual, with no warning.
9. Adding up a shopping bill Adding up has exact rules a programmer can write. Machine learning is for problems too messy for hand-written rules.
10. training data Training data is the system's entire experience of the world.
11. So it sees enough variety to find general patterns Cats come in every colour, pose and lighting. Without variety the system learns something too narrow.
12. Creating an image from a text description Generative AI produces new text, images or sound from patterns in its training data.

Round 2: Quick Fire

1. • It is shown a very large number of labelled examples: its training data
   • It adjusts its internal values to find patterns, getting more labels right
   Model answer

   It is shown hundreds of thousands of images labelled cat or not cat, which is its training data. It adjusts its internal values to find the patterns that go with the label cat, getting more of the labels right each time.

2. • To check it works on new data in the real world
   • Rather than having just memorised its training data
   Model answer

   To check that it has learned general patterns that work on new data in the real world, rather than just memorising the training examples.

3. • Recommendations on streaming or social media
   • Voice assistants and speech to text
   • Face unlock
   • Spam filters
   • Sat nav route prediction
   • Translation
   • Text or image generation
   Model answer

   A spam filter decides which emails are junk and moves them out of the inbox. A streaming service recommends films and shows based on what you have watched before.

4. • It has only found statistical patterns that go with the label
   • So it can be confidently wrong when those patterns mislead it
   Model answer

   It has only found statistical patterns in the pixels that go with the label cat. It has no real understanding, so when those patterns mislead it, it can be confidently and bizarrely wrong.

Round 3: Big Think

1. • Collect a large set of emails labelled spam or not spam
   • Train it: it adjusts its values to find patterns, such as certain words, links or senders
   • Test it on emails it has never seen
   • Use it to filter incoming emails, improving as users mark spam
   Model answer

   First, collect a very large set of emails, each labelled spam or not spam: the training data. Next, train the system, which adjusts its internal values to find patterns that go with spam, such as certain words, suspicious links or unknown senders. Then test it on emails it has never seen, to check it has learned general patterns rather than memorising. Finally, use it to filter incoming emails, and it can keep improving as users mark new messages as spam.

2. • It is fast, giving instant feedback
   • It applies the same rules to every essay
   • Teachers would have more time for teaching
   • It can be confidently wrong, because it does not understand the essay
   • Bias in its training data could penalise unusual or creative writing
   • Students could learn to trick it, for example by stuffing in keywords
   • It is unclear who is responsible for a wrong grade, and how to appeal
   • A reasoned conclusion
   Model answer

   AI marking would be fast, giving students instant feedback, and it would apply the same rules to every essay. It could also reduce teachers' workload, leaving more time for teaching. However, an AI has no real understanding of an essay: it finds patterns, so it can be confidently wrong. If its training data favoured one style of writing, it could be biased against unusual or creative answers, and students might learn to trick it by stuffing in keywords. It is also unclear who is responsible for a wrong grade, or how a student could appeal. On balance, AI could give quick practice feedback, but a teacher should make the final judgement on anything that counts.

This or That

1. Machine learning The system finds the patterns in labelled examples itself, instead of following rules a programmer wrote.
2. AI recognising images Face unlock uses a trained system to recognise your face, one of the everyday uses of AI.
3. Correct labels The system learns from the labels, so wrong labels teach it the wrong patterns.
4. It matches patterns, not the idea of a cat The system has no idea what a cat is. It found statistical patterns that go with the label cat.
5. An everyday use of AI Route prediction learns from enormous amounts of journey data: AI you already meet every day.
6. Its own internal values It adjusts its internal values to get more of the labels right. The images stay the same.
7. Computers doing tasks that need human intelligence AI performs tasks such as recognising objects or understanding language. It does not think or feel.
8. AI understanding language Turning speech into text means recognising language, a classic AI task.

Memory Match

• A 6 Artificial intelligence: Computers doing tasks that need human intelligence
• B 3 Machine learning: Finding patterns in examples instead of following rules
• C 5 Training data: Labelled examples a system learns from
• D 2 Label: The correct answer attached to an example
• E 4 Spam filter: Learns which emails are junk
• F 1 Test data: Unseen examples used to check what it learned