Quiz — Chapter 3: How Machines Learn
Multiple Choice
1. In supervised learning, the term "label" refers to: - (a) The name of the machine learning algorithm being used - (b) The correct answer provided for each training example - (c) The features extracted from the input data - (d) The percentage accuracy of the model
Answer
**(b)** A label is the correct answer or category assigned to each training example. For instance, in a spam detection system, each email in the training set would have a label of "spam" or "not spam."2. Which type of machine learning would be most appropriate for grouping customers into segments based on purchasing behavior, without pre-defined categories? - (a) Supervised learning - (b) Unsupervised learning - (c) Reinforcement learning - (d) Expert system rules
Answer
**(b)** Unsupervised learning is designed to find natural groupings and patterns in data without pre-defined labels or categories. Customer segmentation without predefined groups is a classic unsupervised learning task.3. The purpose of a test set in machine learning is to: - (a) Provide additional examples for the model to learn from - (b) Help tune the model's parameters during development - (c) Evaluate the model's performance on data it has never seen before - (d) Store the model's final predictions for deployment
Answer
**(c)** The test set is held out completely during training and development. It is used only once, at the very end, to evaluate how well the model performs on truly unseen data — providing an honest estimate of how the model will perform in the real world.4. Overfitting occurs when a model: - (a) Fails to learn any patterns from the training data - (b) Learns the training data too well, including noise and quirks, and fails to generalize - (c) Is too simple to capture the real patterns in the data - (d) Uses too little training data
Answer
**(b)** Overfitting happens when a model memorizes the training data rather than extracting general patterns. The telltale sign is high performance on training data but poor performance on new, unseen data.5. In the reinforcement learning framework, an "agent" learns by: - (a) Studying labeled examples provided by human experts - (b) Finding clusters in unlabeled data - (c) Taking actions, receiving rewards or penalties, and adjusting its strategy - (d) Following pre-programmed rules without adaptation
Answer
**(c)** Reinforcement learning involves an agent that interacts with an environment by taking actions, observing the outcomes (rewards or penalties), and gradually adjusting its strategy (policy) to maximize total reward.6. The chapter states that neural networks "learn their own features." This means: - (a) Human programmers specify exactly what features the network should look for - (b) The network discovers what features in the data are important through training - (c) The network copies features from other neural networks - (d) The network uses the same features for every task
Answer
**(b)** Unlike expert systems where humans manually specified what to look for, neural networks discover relevant features automatically during training. In image recognition, for example, a network might learn to detect edges, shapes, and objects without being told these are important.7. Which statement about the relationship between neural networks and the human brain is most accurate? - (a) Neural networks are exact digital replicas of biological brain circuits - (b) Neural networks were inspired by a simplified model of biological neurons but differ fundamentally in mechanism - (c) Neural networks and brains use identical learning processes - (d) There is no relationship between neural networks and the brain
Answer
**(b)** Neural networks were inspired by a simplified model of how biological neurons connect and fire, but the resemblance is superficial. Real neurons are far more complex, the brain's architecture differs fundamentally from artificial neural networks, and biological learning processes are nothing like backpropagation.True or False
8. Supervised learning can only work if the training data is perfectly labeled with no errors.
Answer
**False.** Supervised learning can tolerate some mislabeled examples, though the quality of labels affects the quality of the resulting model. However, systematic biases or large amounts of mislabeling will degrade performance or teach the model incorrect patterns.9. In unsupervised learning, there is always one objectively correct way to cluster the data.
Answer
**False.** Because unsupervised learning has no labels, there is no single "right" answer. Different algorithms or different parameter settings can produce different but equally valid groupings. The appropriateness of a clustering depends on the context and purpose.10. A model that performs poorly on both training data and test data is likely underfitting.
Answer
**True.** Poor performance on both training and test data suggests the model is too simple to capture the real patterns in the data — the hallmark of underfitting.11. The "deep" in deep learning refers to the depth of the model's understanding of its subject matter.
Answer
**False.** "Deep" refers to the many layers in the neural network architecture, not to any metaphorical depth of understanding. A deep neural network has many hidden layers between input and output, allowing it to learn increasingly abstract representations.12. When a machine learning system "learns," it is finding statistical patterns in data rather than developing genuine understanding of a subject.
Answer
**True.** This is the chapter's threshold concept. Machine learning identifies statistical correlations and patterns in training data. While this produces systems that can perform impressively, it is fundamentally different from the understanding, comprehension, and reasoning that characterize human learning.Short Answer
13. Explain why the integrity of the test set is critical in machine learning. What happens if test data accidentally leaks into the training process?
Answer
The test set serves as a final, independent evaluation of the model's ability to generalize to new data. If test data leaks into training (even indirectly), the model may have effectively "seen" the test examples before, inflating its apparent performance. This means the test results no longer provide an honest estimate of how the model will perform on truly new data in the real world. The evaluation becomes meaningless — like a student who has seen the exam questions in advance appearing to perform well without actually learning the material.14. In your own words, explain what "generalization" means in machine learning and why it matters.
Answer
Generalization is a model's ability to perform well on new, previously unseen data — not just the data it was trained on. It matters because the entire point of machine learning is to learn patterns that apply broadly, not just to memorize specific examples. A model that generalizes well has captured genuine, underlying patterns in the data. A model that fails to generalize has merely memorized the training data (overfitting) and will be unreliable when deployed in the real world.15. Describe the reinforcement learning loop in four steps, using a concrete example of your choosing (not one from the chapter).
Answer
[Answers will vary. Example using a thermostat-like system learning to manage building temperature:] 1. **Observe:** The system measures the current room temperature (68°F) and outside temperature (35°F). 2. **Choose an action:** Based on its current strategy, it increases heating by 5%. 3. **Receive feedback:** The room reaches the target temperature of 72°F efficiently (positive reward) without excessive energy use (no penalty). 4. **Update strategy:** The system adjusts its internal strategy to favor this level of heating increase under similar conditions in the future. This cycle repeats continuously, refining the strategy over time.Applied Scenario
16. A company develops an AI tool that screens resumes for a tech firm. The system is trained on 10 years of the company's hiring data — using the resumes of people who were hired as "positive" examples and those who weren't as "negative" examples. The system achieves 89% accuracy on the test set. Analyze this scenario by addressing: (a) What type of machine learning is being used? (b) What potential biases might exist in the training data? (c) Is 89% accuracy necessarily good? (d) What questions would you ask before trusting this system?
Answer
**(a)** Supervised learning — the system learns from labeled examples (hired = positive, not hired = negative) to predict labels for new resumes. **(b)** Potential biases include: historical biases in hiring decisions (e.g., if the company historically preferred candidates from certain schools, genders, or backgrounds, the system will learn those preferences as "what a good candidate looks like"); survivorship bias (the training data only reflects who was hired, not who would have been successful if given the chance); changing job requirements (what made a good hire 10 years ago may not reflect current needs). **(c)** 89% accuracy is not necessarily good. We need to know: What's the baseline? (If 85% of applicants are rejected, a system that rejects everyone would be "85% accurate.") What kinds of errors does it make? (Rejecting a qualified candidate versus advancing an unqualified one have different consequences.) How does 89% compare to human screeners? **(d)** Questions include: Was the test set representative of the actual applicant pool? Has the system been tested for disparate impact across demographic groups? What features is the system using — could it be using proxies for protected characteristics (zip code as proxy for race, name as proxy for gender)? What happens to the 11% it gets "wrong"?17. Priya uses a generative AI tool to help write an essay on the causes of the French Revolution. The tool produces a well-structured, fluent essay that includes a quote it attributes to historian Simon Schama. Priya looks up the quote and discovers that while Schama did write about the French Revolution, this specific quote doesn't appear in any of his works. Using concepts from this chapter, explain: (a) Why the AI tool produced a fabricated quote, (b) Why the fabricated quote "sounds right," and (c) What this tells us about the threshold concept from Section 3.7.