The worksheetDownload the PDF
Answers

F12.7 AI, robots and bias

Technology and society · GCSE · OCR J277 1.6.1, AQA 8525 3.8, Edexcel 1CP2 5.2.2 · about 20 min

BugBotLab

What this lesson is about

What AI is good for, where it goes wrong, and auditing a training set for bias.

Questions 6 marks in all

  1. [1 mark]What is machine learning?

    1. AA program that finds patterns from examples instead of being given the rules
    2. BA robot that moves by itself
    3. CA faster processor
    4. DA kind of database
    Answer: A. The rules are learned from the training data.

  2. [1 mark]Why can a face recognition system work worse for some groups?

    1. AIts training data did not represent them well
    2. BTheir faces are harder to see
    3. CCameras are biased by design
    4. DThe software is too fast
    Answer: A. A model learns from the data it is shown.

  3. [1 mark]Which is a real difficulty with machine learning decisions?

    1. AIt can be hard to explain why a decision was made
    2. BThey are always wrong
    3. CThey cannot use data
    4. DThey need no electricity
    Answer: A. The rules are learned, not written, so they cannot simply be read.

  4. [1 mark]A self-driving car injures someone. What is the hard question?

    1. AWho is accountable: owner, manufacturer or programmer
    2. BWhether the car was fast
    3. CWhat colour the car was
    4. DWhich fuel it used
    Answer: A. Accountability for automated decisions is still being settled in law.

  5. [1 mark]A training set has 20 samples and 3 are 'night'. What percentage is that?

    Answer: 15. 3 ÷ 20 × 100.

  6. [1 mark]What does this program print?

    counts = {}
for s in ['day', 'day', 'night']:
    counts[s] = counts.get(s, 0) + 1
print(counts)
    Answer:
    {'day': 2, 'night': 1}

    Two days and one night.

The task: audit the training set

Count how many samples of each label are in samples. Print <label>: <n> (<share>%) for each, sorted from the most common to the least, with the share rounded to the nearest whole number. Any label making up less than 15% is under-represented: add UNDER-REPRESENTED to its line. At the end print labels: <n> and under-represented: <labels>, the under-represented labels joined by , .

# the two lines every program starts with: the commands, then the robot
from bugbot import *
connect()

samples = [
    "day", "day", "day", "day", "day", "day", "day", "day", "day", "day",
    "day", "day", "day", "dusk", "dusk", "night", "night", "night", "indoor", "day",
]

The hint students can ask for: Count each label into a dictionary, then sort from the most common down. A label's share is its count out of the total; mark the ones below the threshold and collect them for the last line.

A solution

from bugbot import *
connect()
samples = [
    "day", "day", "day", "day", "day", "day", "day", "day", "day", "day",
    "day", "day", "day", "dusk", "dusk", "night", "night", "night", "indoor", "day",
]
counts = {}
for s in samples:
    counts[s] = counts.get(s, 0) + 1
under = []
for label, n in sorted(counts.items(), key=lambda kv: kv[1], reverse=True):
    share = n / len(samples) * 100
    if share < 15:
        under.append(label)
        print(f"{label}: {n} ({share:.0f}%) UNDER-REPRESENTED")
    else:
        print(f"{label}: {n} ({share:.0f}%)")
print("labels:", len(counts))
print("under-represented:", ", ".join(under))

Any program that meets the task's checks is marked correct in the simulator; this is one way, not the only way.