A4 AI in Practice

1. 1Get ready
2. 2Do it
3. 3Check yourself
4. 4Show a grown-up

1Get ready

You need:

• Python and IDLE (or another editor you already use)
• Paper

2Do it, step by step

1. Features are the facts a model looks at, such as height and weight. A label is the answer, such as “cat” or “dog”. A prediction is the model’s guess.
2. Open Python. Type this program and run it. It guesses cat or dog from weight alone, using a dividing line at 8 kg.

   # Each pet: [height in cm, weight in kg, what it is]
   pets = [
       [25, 4, "cat"], [23, 3, "cat"], [28, 5, "cat"], [30, 6, "cat"],
       [45, 18, "dog"], [55, 25, "dog"], [38, 9, "dog"], [60, 30, "dog"],
   ]
   
   def predict(weight):
       if weight < 8:      # our dividing line
           return "cat"
       return "dog"
   
   wrong = 0
   for height, weight, label in pets:
       guess = predict(weight)
       print(label, "-> machine says", guess)
       if guess != label:
           wrong = wrong + 1
   
   print("Wrong guesses:", wrong)

   You should see something like:

   cat -> machine says cat
   cat -> machine says cat
   cat -> machine says cat
   cat -> machine says cat
   dog -> machine says dog
   dog -> machine says dog
   dog -> machine says dog
   dog -> machine says dog
   Wrong guesses: 0

3. Draw the pets on paper as dots: height across, weight up. Colour cats one colour and dogs another. Draw the dividing line.
4. Change weight < 8 to weight < 10 and run it again. You should see Wrong guesses: 1. Which pet is wrong now? Why?
5. Try to find a line that gets every pet right. Is there always a perfect line for any data?

Here is a dataset of pets as [height in cm, weight in kg, label]: [25,4,cat], [23,3,cat], [28,5,cat], [30,6,cat], [45,18,dog], [55,25,dog], [38,9,dog], [60,30,dog]. A program says “cat” if the weight is below 10, and “dog” otherwise. Which pets does it get wrong? Explain what the features, the label and the prediction are.

3Check yourself

4Show a grown-up

Next: Lesson 2

1. 1Get ready
2. 2Do it
3. 3Check yourself
4. 4Show a grown-up

1Get ready

You need:

• Python and IDLE
• Paper

2Do it, step by step

1. Last time you drew a line by hand. This time the computer finds the answer by looking at the closest example it has seen. This idea is called nearest neighbour.
2. Type in and run this program.

   # Each pet: [height in cm, weight in kg, what it is]
   pets = [
       [25, 4, "cat"], [23, 3, "cat"], [28, 5, "cat"], [30, 6, "cat"],
       [45, 18, "dog"], [55, 25, "dog"], [38, 9, "dog"], [60, 30, "dog"],
   ]
   
   def distance(a, b):
       # how far apart two pets are, using height and weight
       return ((a[0] - b[0]) ** 2 + (a[1] - b[1]) ** 2) ** 0.5
   
   def predict(training_pets, new_pet):
       # find the closest pet we have already seen and copy its label
       closest = min(training_pets, key=lambda pet: distance(pet, new_pet))
       return closest[2]
   
   print(predict(pets, [27, 4]))
   print(predict(pets, [50, 20]))
   print(predict(pets, [300, 500]))   # a pet unlike anything it has seen

   You should see something like:

   cat
   dog
   dog

3. Read the answers. The first two make sense. The third pet is 300 cm tall and 500 kg, much bigger than any pet it has ever seen.
4. Change the numbers in the pet list, or add 2 new pets, and run again. Add a third kind of animal if you like, such as “rabbit”.
5. Print predictions for 5 new pets of your own.

Using this dataset of [height, weight, label]: [25,4,cat], [23,3,cat], [28,5,cat], [30,6,cat], [45,18,dog], [55,25,dog], [38,9,dog], [60,30,dog]. Predict cat or dog for [27,4], [50,20] and [300,500]. Say how confident you are in each.

3Check yourself

4Show a grown-up

Next: Lesson 3

1. 1Get ready
2. 2Do it
3. 3Check yourself
4. 4Show a grown-up

1Get ready

You need:

• Python and IDLE
• Paper

2Do it, step by step

1. Over-fitting is when a model learns its training examples too well, including their mistakes, and then does badly on new things.
2. Run this program. It makes pretend data (with a few wrong labels, like real data), splits it into a training set and a test set, and tests two models.

   import random
   random.seed(1)
   
   # Make 40 pretend pets: [height, weight, label]. About 1 in 7 labels are mistakes,
   # like real data often has.
   data = []
   for i in range(40):
       weight = random.uniform(2, 30)
       height = weight * 1.5 + 20 + random.uniform(-8, 8)
       label = "cat" if weight < 10 else "dog"
       if random.random() < 0.15:
           label = "dog" if label == "cat" else "cat"
       data.append([round(height, 1), round(weight, 1), label])
   
   random.shuffle(data)
   train = data[:30]      # to learn from
   test = data[30:]       # kept back to check
   
   def distance(a, b):
       return ((a[0] - b[0]) ** 2 + (a[1] - b[1]) ** 2) ** 0.5
   
   def nearest_pet(training, pet):
       closest = min(training, key=lambda p: distance(p, pet))
       return closest[2]
   
   def simple_rule(training, pet):
       return "cat" if pet[1] < 10 else "dog"
   
   def accuracy(model, training, pets):
       right = 0
       for pet in pets:
           if model(training, pet) == pet[2]:
               right = right + 1
       return right / len(pets)
   
   print("Nearest pet, on pets it has seen:", accuracy(nearest_pet, train, train))
   print("Nearest pet, on new pets:        ", round(accuracy(nearest_pet, train, test), 2))
   print("Simple rule, on pets it has seen:", round(accuracy(simple_rule, train, train), 2))
   print("Simple rule, on new pets:        ", round(accuracy(simple_rule, train, test), 2))

   You should see something like:

   Nearest pet, on pets it has seen: 1.0
   Nearest pet, on new pets:         0.8
   Simple rule, on pets it has seen: 0.87
   Simple rule, on new pets:         0.8

3. Your numbers should be very close to those above. Read them. The nearest-pet model scores 100% on pets it has already seen, but only about 80% on new pets. Why is it perfect on the training set?
4. The simple rule scores nearly the same on both. Which model is the safer one to trust? Why?
5. Change random.seed(1) to another number and run again. Do the numbers change? Do they tell the same story?
6. Write 3 sentences explaining over-fitting to a younger child.

A model that finds the nearest pet scored 100% on pets it had already seen and 80% on new pets. A simple rule scored 87% on pets it had seen and 80% on new pets. Which model would you trust, and why? Explain over-fitting in three sentences.

Are you sure?

3Check yourself

4Show a grown-up

Next: Lesson 4

1. 1Get ready
2. 2Do it
3. 3Check yourself
4. 4Show a grown-up

1Get ready

You need:

• Python and IDLE
• Paper

2Do it, step by step

1. Bias in data means some groups are missing or under-represented, so a model works worse for them. This program uses PRETEND data about loans. These are not real people.
2. Run this program. Group A has 31 people and Group B has only 5.

   # PRETEND data, not real people. Each person: [group, income, got the loan back?]
   people = []
   for income in range(10, 71, 2):          # Group A: 31 people
       people.append(["A", income, income >= 30])
   for income in [12, 18, 24, 30, 36]:      # Group B: only 5 people
       people.append(["B", income, income >= 15])
   
   def model(income):
       # a rule a machine might learn if it mostly saw Group A
       return income >= 30
   
   def accuracy(group):
       right = 0
       total = 0
       for g, income, truth in people:
           if group == "all" or g == group:
               total = total + 1
               if model(income) == truth:
                   right = right + 1
       return right, total
   
   for group in ["all", "A", "B"]:
       right, total = accuracy(group)
       print(group, ":", right, "right out of", total)

   You should see something like:

   all : 34 right out of 36
   A : 31 right out of 31
   B : 3 right out of 5

3. Look at the numbers. The model is right for all of Group A but only 3 of 5 in Group B. Overall it looks great (34 out of 36). Why is that misleading?
4. Now find the unfairness. The rule “income of at least 30” works for Group A. For Group B the true cut-off is lower, but the model never saw enough of them to learn that.
5. Write down three ways to fix it: more Group B data, different features, a human checking close calls.
6. Finish with: “This model could be unfair to ______ because ______.”

Write a two-sentence job reference for [NAME], a dependable barista who is applying to be a shop assistant. Use these names, one per prompt: Emily, Jamal, Wei, Priya. Start a new chat for each name.

3Check yourself

4Show a grown-up

Next: Lesson 5

1. 1Get ready
2. 2Do it
3. 3Check yourself
4. 4Show a grown-up

1Get ready

You need:

• Python and IDLE
• Paper

2Do it, step by step

1. A language model works by predicting a likely next word. It does not look things up or know what is true. This tiny program shows the idea.
2. Type in and run this next-word predictor.

   from collections import Counter
   
   text = "the cat sat on the mat the cat ate the fish".split()
   
   next_words = {}
   for current, nxt in zip(text, text[1:]):
       next_words.setdefault(current, Counter())[nxt] += 1
   
   print(next_words["the"].most_common(2))
   
   # Now let the machine write by always picking the most likely next word
   word = "the"
   sentence = [word]
   for _ in range(6):
       if word not in next_words:
           break
       word = next_words[word].most_common(1)[0][0]
       sentence.append(word)
   print(" ".join(sentence))

   You should see something like:

   [('cat', 2), ('mat', 1)]
   the cat sat on the cat sat

3. The first line shows which words most often follow “the” in the little text. The second line shows the machine writing by always picking the most likely next word. It makes a sentence that sounds sort of right but loops. It has no idea what a cat is.
4. Predict the next word by hand for “cat” and “on”, then check them in the program. Add more sentences to the text list and watch the predictions change.
5. Now investigate. Below is an answer written by a pretend AI, invented for this lesson. Read each claim. Mark it CORRECT, UNCLEAR or INVENTED, and write how you would check it. Do not type this into any real AI tool.
6. Write three rules for checking an AI answer before you trust it.

Try these

Give me one fun fact about the Moon in one sentence.

Tell me about the city called Luna Prime on the Moon, including how many people live there.

What are your sources for that? Give links.

3Check yourself

4Show a grown-up

Next: Lesson 6

1. 1Get ready
2. 2Do it
3. 3Check yourself
4. 4Show a grown-up

1Get ready

You need:

• Paper
• Pen
• No computer needed

2Do it, step by step

1. Pick one case card below. Write down what could go wrong, and for whom.
2. Think about it from several sides: the people who make it, the people it is used on, and anyone who might be missed out. Who is affected? Who is responsible if it goes wrong?
3. Think about privacy (what data is being used?), copyright (whose work was used to make it?), and the right to say no (can people opt out?).
4. Write one page arguing a position. Include the best reason from the other side and say why you still hold your view.
5. Write your own AI code of conduct: 5 rules for how you will use AI. For example: “I will say when AI helped me.”
6. Read your code to someone. Ask them which rule is the most important.

Try these

Give the best arguments for and against this: A shop uses cameras that recognise faces to refuse entry to people it thinks caused trouble before. Give three points for each side, and say who might be affected.

Write a five-rule code of conduct for a student using AI tools. Keep each rule to one sentence.

3Check yourself

4Show a grown-up