Week 01 · Page 4 of 4 · Self-Check Quiz

Ten questions. If you miss more than two, return to the Lesson page before starting Week 2 — everything that follows assumes this vocabulary is solid.

Quiz Self-Check 10 Questions

Answer each question yourself before revealing the answer. Explanations are provided for every question, including ones you may answer correctly, since the reasoning matters more than the score.

Question 1

Which of the following best defines machine learning, as used in this course?

  1. A method for deriving physical laws from first principles
  2. A method for building a predictive model by fitting it to examples of data
  3. Any computer program that processes scientific data
  4. A technique exclusive to neural networks
Show answer
Correct answer: B. Machine learning, in this course's working definition, means fitting a model's parameters to examples rather than deriving the model from theory. Option A describes traditional physics modeling; C is too broad (it would include any simulation code); D wrongly restricts the term to one specific method.
Question 2

In the sentence “we trained a model on features to predict a label,” which part is the label?

  1. The model itself
  2. The input quantities describing each example
  3. The known output value used during training
  4. The training algorithm
Show answer
Correct answer: C. The label (or target) is the known output value attached to each training example. Features are the inputs (option B); the model is the object making predictions (A); the training algorithm is the procedure that adjusts the model's parameters (D).
Question 3

A dataset has no known property values for any of its entries, only computed structural descriptors. This is:

  1. A supervised learning problem
  2. An unsupervised learning problem
  3. Not a machine learning problem at all
  4. Impossible to determine without more information
Show answer
Correct answer: B. No labels are present — only features (the structural descriptors). By definition, this is an unsupervised setting; any structure discovered (e.g. clusters) would be unsupervised learning.
Question 4

You train a model on formation energies for compounds with unit-cell volumes between 8 and 16 ų/atom. Why should you be cautious about predicting on a compound with volume 25 ų/atom?

  1. The model will definitely fail
  2. This is extrapolation outside the training range, where the fitted relationship is not guaranteed to hold
  3. Models can only be applied to volumes that are exact multiples of training values
  4. There is no reason for caution if the model fit the training data well
Show answer
Correct answer: B. A model is only empirically validated within (or close to) the range of data used to train it. Predicting far outside that range is extrapolation, and even an excellent fit on the training data is no guarantee the same relationship continues to hold outside it.
Question 5

What distinguishes a feature from a label?

  1. Features are always integers, labels are always real numbers
  2. A feature is an input used to make a prediction; a label is the known output the model is trained to predict
  3. There is no real distinction — the terms are interchangeable
  4. Labels are only used in unsupervised learning
Show answer
Correct answer: B. A feature is an input describing an example; a label is the known correct output for that example, used during supervised training. Labels do not exist in unsupervised learning (option D is backwards), and there's no rule about integers vs. real numbers (option A).
Question 6

In Worked Example A (formation energy vs. electronegativity difference), what did ‘training the model’ actually consist of?

  1. Running a new DFT calculation for each compound
  2. Finding numerical values for the slope and intercept of a line that fit the known data points well
  3. Measuring the electronegativity of each element experimentally
  4. Choosing which compound to predict next
Show answer
Correct answer: B. Training meant finding parameter values (here, the slope and intercept of a linear model) that made the model's predictions match the known labels as closely as possible — exactly the general definition of training given in the Lesson.
Question 7

A materials scientist has 1,200 compounds with computed band gaps and wants a model to predict band gap for new compositions. What kind of problem is this, and why?

  1. Unsupervised, because there are many compounds
  2. Supervised, because every compound has a known label (the band gap)
  3. Neither, because band gap is a continuous quantity
  4. Cannot be determined without knowing the features
Show answer
Correct answer: B. The presence of a known label (band gap) for every training example is what makes this supervised, regardless of how many compounds there are or whether the label is continuous (regression) or categorical (classification).
Question 8

Why does the Lesson page insist that unsupervised learning is not just ‘supervised learning without enough data yet’?

  1. Because unsupervised methods always need less data than supervised ones
  2. Because the two settings answer fundamentally different questions, even with infinite data — prediction of a label vs. discovery of structure
  3. Because unsupervised learning is an older technique
  4. Because supervised learning cannot be used on small datasets
Show answer
Correct answer: B. Even with unlimited data, supervised learning answers ‘what is the label of this new example’ while unsupervised learning answers ‘what structure exists among these examples’ — a question that remains meaningful even when no label is ever defined for the data.
Question 9

In Worked Example B, what was the actual evidence for there being two groups of compounds?

  1. A computed property that confirmed two distinct categories
  2. The visual clustering of points when plotted by their two features
  3. A label provided in the original dataset
  4. A statistical hypothesis test
Show answer
Correct answer: B. No label or formal test was used; the apparent grouping was read directly off a plot of the two features. The Lesson explicitly notes this is a hypothesis about structure, not a proven physical fact, until checked against independent evidence.
Question 10

What is the main skill the Week 1 Practical Work is designed to train?

  1. Writing scikit-learn code
  2. Correctly diagnosing whether a real research problem is supervised or unsupervised, and identifying plausible features, before any algorithm is chosen
  3. Memorizing the names of machine learning algorithms
  4. Performing a DFT calculation
Show answer
Correct answer: B. Week 1 deliberately has no coding. The practical trains problem diagnosis — correctly identifying the learning setting and candidate features for a real problem — which the Practical Work page states explicitly is a skill used more often in real research than any specific algorithm.
Scoring guide

9–10 correct: you are ready for Week 2. 6–8 correct: re-read the definition boxes on the Lesson page for any term you missed. 5 or fewer: re-read the full Lesson page before attempting Week 2, as the statistical concepts introduced next week (bias, variance, cross-validation) are defined directly in terms of this week's vocabulary.

← Practical Work Quiz · Page 4 of 4 Week 2 →

INTELLIGENT PHYSICS: MACHINE LEARNING METHODS FOR MATERIALS SCIENCE — WEEK 1, QUIZ