MskProdComputing
Back to the topic

AI · 2.2

How Machines Learn: Supervised, Unsupervised and Reinforcement

Section A: 10 multiple choice questions, 10 marks. Section B: 3 written questions, 11 marks. Total 21 marks.

Name

Class

Date

Mark  / 21

Section A

Circle the letter of the correct answer. One mark each.

  1. 1In Mitchell's definition, what is P for a spam filter?1
    1. AThe percentage of emails it sorts correctly
    2. BThe emails people have labelled
    3. CSorting emails into spam and not spam
    4. DThe number of emails received
  2. 2What is inference?1
    1. AAdjusting the parameters to fit the examples
    2. BDeleting a model after use
    3. CUsing a trained model to produce outputs for new inputs
    4. DCollecting and labelling the training data
  3. 3Predicting the price of a house from its size and location is an example of:1
    1. Aclustering
    2. Breinforcement learning
    3. Cclassification
    4. Dregression
  4. 4Which is unsupervised learning?1
    1. ATeaching a robot to walk with rewards
    2. BGrouping customers into segments by their shopping habits
    3. CSorting X-rays that doctors have labelled
    4. DPredicting a house price from labelled sales
  5. 5Why are delayed rewards a problem in reinforcement learning?1
    1. AIt is hard to tell which earlier actions deserve the credit
    2. BThe environment stops while it waits
    3. CThe agent cannot take actions until the reward arrives
    4. DDelayed rewards are always negative
  6. 6How does self-supervised learning get its labels?1
    1. ABy copying labels from another model
    2. BBy hiding part of the data and using the hidden part as the label
    3. CBy paying people to label every example
    4. DBy rewarding the model for good answers
  7. 7In machine learning, what is a feature?1
    1. AA measurable property of an example used as input
    2. BThe model's final prediction
    3. CThe correct answer for an example
    4. DA bug in the model
  8. 8What is transfer learning?1
    1. ACopying training data from one computer to another
    2. BAdapting a model trained on a large general task to a specific task
    3. CTraining two models at the same time
    4. DMoving a model from training to inference
  9. 9A model learns to predict whether a photo shows a cat or a dog from 50,000 labelled photos. Which kind of learning is this?1
    1. ASupervised regression
    2. BUnsupervised clustering
    3. CReinforcement learning
    4. DSupervised classification
  10. 10What does a nearest neighbour classifier do with a new input?1
    1. AAsks a person for the label
    2. BAverages all the labels in the training data
    3. CGives it the label of the most similar training example
    4. DPicks a label at random

Section B

Answer in the space provided. The number of marks tells you how many separate points to make.

  1. 1ApplyApply Mitchell's definition of learning to a program that learns to recommend films.3
  2. 2CompareCompare supervised and unsupervised learning.4
  3. 3ExplainExplain why self-supervised learning made large language models possible.4

Mark scheme

Section A

1A2C3D4B5A6B7A8B9D10C

Section B

1 Apply Mitchell's definition of learning to a program that learns to recommend films. 3

One mark for each of the following, to a maximum of 3.

  • Task: recommending films a person will like
  • Performance: e.g. percentage of recommendations watched or rated highly
  • Experience: records of what people watched or rated

Model answer. The task T is recommending films that a person will enjoy. The performance measure P could be the percentage of recommended films that people go on to watch, or rate highly. The experience E is the record of what people have watched, finished, skipped and rated. The program learns if its recommendations are watched more often as that record grows.

2 Compare supervised and unsupervised learning. 4

One mark for each of the following, to a maximum of 4.

  • Supervised learning uses labelled examples / correct answers
  • Unsupervised learning has no labels
  • Supervised predicts a known output, such as a class or number
  • Unsupervised finds structure such as clusters or anomalies

Model answer. Supervised learning trains on examples that include the correct answer, the label, so the model learns to predict a known kind of output, either a category in classification or a number in regression. Unsupervised learning has no labels, so instead of predicting answers it looks for structure in the inputs, such as clusters of similar examples or unusual anomalies. Supervised learning needs costly labelling, while unsupervised results need a person to judge whether the structure found is meaningful.

3 Explain why self-supervised learning made large language models possible. 4

One mark for each of the following, to a maximum of 4.

  • Labelling trillions of words by hand would be impossible / too costly
  • Self-supervised learning hides part of the text and predicts it
  • So the labels come free from the text itself
  • Predicting text well requires learning grammar, facts or patterns

Model answer. Large language models learn from trillions of words, and paying people to label that much text would be impossible. Self-supervised learning avoids labelling altogether: the model is shown text with the next word hidden and learns to predict it, so every word in the text is a free label. To predict the next word well across so much varied text, the model has to pick up grammar, facts and patterns of reasoning, which is what makes it useful.