Unit 6: Machine Learning - Practice Quiz

ECAP792 60 Questions
0 Correct 0 Wrong 60 Left
0/60

1 What is machine learning?

Introduction Easy
A. A method that enables computers to learn from data
B. A method for manually entering data
C. A method for connecting computers to networks
D. A method for increasing computer storage

2 What is the main purpose of training a machine learning model?

Introduction Easy
A. To remove all data from a system
B. To replace the operating system of a computer
C. To learn patterns that support predictions
D. To increase the physical memory of a system

3 In machine learning, what is a model?

Introduction Easy
A. A learned representation used to make predictions
B. A device used to collect physical data
C. A folder used to store computer programs
D. A network used to connect databases

4 What is training data?

Introduction Easy
A. Data used only to install software
B. Data used only to design a website
C. Data used only to format a disk
D. Data used to teach a machine learning model

5 Which machine learning technique learns from labeled examples?

Types of machine learning techniques Easy
A. Reinforcement learning
B. Random learning
C. Supervised learning
D. Unsupervised learning

6 Which technique is commonly used to discover groups in unlabeled data?

Types of machine learning techniques Easy
A. Reinforcement learning
B. Supervised learning
C. Transfer learning
D. Unsupervised learning

7 Which technique learns through rewards and penalties?

Types of machine learning techniques Easy
A. Batch learning
B. Supervised learning
C. Unsupervised learning
D. Reinforcement learning

8 Which technique uses both labeled and unlabeled training examples?

Types of machine learning techniques Easy
A. Reinforcement learning
B. Semi-supervised learning
C. Supervised learning
D. Unsupervised learning

9 Which learning problem predicts a category such as "spam" or "not spam"?

Learning problems and system Easy
A. Dimensionality reduction
B. Classification
C. Clustering
D. Regression

10 Which learning problem predicts a continuous value such as house price?

Learning problems and system Easy
A. Classification
B. Regression
C. Association
D. Clustering

11 Which learning problem places similar data points into groups?

Learning problems and system Easy
A. Classification
B. Ranking
C. Regression
D. Clustering

12 What is overfitting in a machine learning system?

Learning problems and system Easy
A. The model uses too few input features
B. The model learns training data but performs poorly on new data
C. The model stores too little training data
D. The model trains for exactly one cycle

13 What should usually be identified first when designing a learning system?

Designing a learning system Easy
A. The file name
B. The learning task
C. The processor brand
D. The screen color

14 Why is relevant training data collected when designing a learning system?

Designing a learning system Easy
A. To teach the model useful patterns
B. To change the computer operating system
C. To increase the monitor resolution
D. To rename the project files

15 What is a feature in a machine learning system?

Designing a learning system Easy
A. A backup copy of the trained program
B. An input attribute used by the model
C. A final score assigned after testing
D. A hardware device connected to a computer

16 Why is test data used after training a model?

Designing a learning system Easy
A. To select the computer's operating system
B. To add labels to every training example
C. To evaluate performance on unseen examples
D. To increase the size of every feature

17 In the definition of a learning problem, what does the task describe?

Concept of learning task Easy
A. The job the system must perform
B. The language used for documentation
C. The number of files in a folder
D. The computer used to store data

18 In a learning task, what does experience refer to?

Concept of learning task Easy
A. The hardware selected for deployment
B. The color chosen for the interface
C. The folder containing the application
D. The data or interactions used for learning

19 What is the purpose of a performance measure in a learning task?

Concept of learning task Easy
A. To define the format of every file
B. To evaluate how well the system performs
C. To count the number of installed programs
D. To choose the physical size of a computer

20 For an email spam detector, what is the learning task?

Concept of learning task Easy
A. Measuring the computer's storage capacity
B. Classifying emails as spam or not spam
C. Grouping computers by processor speed
D. Predicting the length of an email address

21 A model achieves 98% accuracy on its training data but only 72% accuracy on previously unseen data. Which conclusion is most appropriate?

Introduction Medium
A. The model has eliminated prediction variance
B. The model is underfitting the training data
C. The model is overfitting the training data
D. The model requires no further evaluation

22 A dataset contains house area, number of rooms, neighborhood, and selling price. If the goal is to predict selling price, how should these variables be used?

Introduction Medium
A. All variables are targets with no input features
B. House area is the target; the others are features
C. Selling price is the target; the others are features
D. Neighborhood is the target; the others are features

23 A bank trains a model using transactions labeled as either fraudulent or legitimate. Which learning technique is being used?

Types of machine learning techniques Medium
A. Reinforcement learning
B. Supervised learning
C. Association learning
D. Unsupervised learning

24 A retailer wants to divide customers into groups based on purchasing behavior, but no customer-group labels are available. Which technique is most suitable?

Types of machine learning techniques Medium
A. Classification
B. Reinforcement learning
C. Regression
D. Clustering

25 A warehouse robot improves its navigation by receiving positive rewards for fast deliveries and penalties for collisions. Which technique best describes this process?

Types of machine learning techniques Medium
A. Linear regression
B. Unsupervised clustering
C. Reinforcement learning
D. Supervised classification

26 A company has 500 labeled medical images and 20,000 unlabeled images. It wants to use both sets to train a classifier. Which approach is most appropriate?

Types of machine learning techniques Medium
A. Purely unsupervised learning
B. Semi-supervised learning
C. Reinforcement learning
D. Purely supervised learning

27 Which pairing correctly matches a prediction problem with its supervised learning type?

Types of machine learning techniques Medium
A. Discovering customer segments: classification
B. Grouping news articles: regression
C. Predicting loan default: classification
D. Predicting annual rainfall: classification

28 A polynomial model performs poorly on both its training set and test set. What is the most likely learning problem?

Learning problems and system Medium
A. Data leakage
B. Underfitting
C. Overfitting
D. Concept drift

29 A model trained to predict employee resignations uses a feature indicating whether an exit interview was completed. Why is this feature problematic?

Learning problems and system Medium
A. It creates unsupervised labels
B. It causes class balancing
C. It reduces feature dimensionality
D. It introduces target leakage

30 An email classifier worked well when deployed, but its accuracy gradually declined as users adopted new vocabulary and writing styles. What is the most likely cause?

Learning problems and system Medium
A. Concept drift
B. Feature scaling
C. Random initialization
D. Label encoding

31 Two classes contain 9,900 legitimate transactions and 100 fraudulent transactions. Why can 99% accuracy be misleading?

Learning problems and system Medium
A. Fraud labels become continuous values
B. Training examples lose their features
C. Accuracy always ignores correct predictions
D. A majority-only model can achieve it

32 A team must choose among several model configurations. Which dataset allocation best supports unbiased final evaluation?

Designing a learning system Medium
A. Train and tune on test data, then evaluate on validation data
B. Train and evaluate on all data, then create a test set
C. Train on validation data, tune on test data, and evaluate on training data
D. Train on training data, tune on validation data, and evaluate on test data

33 For a disease-screening system, failing to identify an affected patient is much more costly than issuing an extra warning. Which metric should receive the greatest emphasis?

Designing a learning system Medium
A. Compression ratio
B. Training speed
C. Specificity
D. Recall

34 A speech-recognition model will serve users in many regions. Which training-data strategy is most appropriate?

Designing a learning system Medium
A. Use only studio recordings from one frequently represented region
B. Select recordings that produce the highest training accuracy
C. Sample speakers, accents, and environments expected after deployment
D. Exclude uncommon accents to make the dataset more consistent

35 A linear model cannot capture a curved relationship between two variables. Which design change most directly addresses this limitation?

Designing a learning system Medium
A. Evaluate repeatedly on the same test observations
B. Introduce nonlinear features or use a nonlinear model
C. Increase the classification threshold above its maximum
D. Remove the target values from the training dataset

36 A recommendation system is retrained every week from user clicks. Which precaution best limits a feedback loop in which already-popular items dominate future recommendations?

Designing a learning system Medium
A. Collect some interaction data through controlled exploration
B. Remove all records involving less-popular products
C. Train only on items shown by the current model
D. Rank products solely by their historical click totals

37 Using the task-experience-performance framework, a chess program learns from completed games and is evaluated by the percentage of games won. What represents the experience ?

Concept of learning task Medium
A. The available legal moves
B. The act of playing chess
C. The percentage won
D. The completed games

38 A spam filter learns from previously labeled emails and is judged by its score on new emails. What is the learning task ?

Concept of learning task Medium
A. Reading the collection of previously labeled emails
B. Computing the harmonic mean of precision and recall
C. Storing incoming messages in the user's mailbox
D. Classifying new emails as spam or non-spam

39 A delivery-time model predicts the number of minutes required for each order. Which formulation correctly identifies its output and task type?

Concept of learning task Medium
A. A continuous output in a regression task
B. A discrete output in a classification task
C. A reward sequence in a reinforcement task
D. An unlabeled group in a clustering task

40 A model minimizes mean squared error, . Which change has the greatest direct effect on this measure?

Concept of learning task Medium
A. A prediction with a very small error
B. A label written as an integer
C. A feature with a binary value
D. A prediction with a very large error

41 Two classifiers have identical zero-one loss on a finite training set but make different predictions on unseen inputs. Which principle best explains why empirical evidence alone cannot determine which classifier will generalize better?

Introduction Hard
A. The central limit theorem requires normally distributed labels
B. The no-free-lunch principle requires an inductive bias
C. The law of large numbers requires zero training error
D. The representer theorem requires a linear hypothesis class

42 Suppose labels satisfy , where and . Under squared-error loss, what is the lowest achievable expected prediction error for any function of ?

Introduction Hard
A. , because inputs vary across examples
B. , because conditional noise remains unpredictable
C. , because the Bayes predictor recovers every label
D. , because error is measured in label units

43 A high-capacity model's training error decreases after more features are added, while its test error increases. Assuming the evaluation data are representative and uncontaminated, which diagnosis is most defensible?

Introduction Hard
A. The added features increased estimation variance
B. The added features corrected label imbalance
C. The added features increased approximation bias
D. The added features eliminated irreducible noise

44 For binary classification under equal misclassification costs, a calibrated model estimates . Which prediction minimizes conditional zero-one risk at , and what is that risk?

Introduction Hard
A. Predict ; the conditional risk is
B. Predict ; the conditional risk is
C. Predict ; the conditional risk is
D. Predict ; the conditional risk is

45 A model learns representations by predicting masked words in millions of unlabeled documents, then is fine-tuned using labeled sentiment examples. How should the two phases be classified?

Types of machine learning techniques Hard
A. Unsupervised clustering followed by active learning
B. Supervised learning followed by transfer-free inference
C. Reinforcement learning followed by semi-supervised learning
D. Self-supervised learning followed by supervised learning

46 A recommendation system selects one item, observes a reward only for that item, and does not model long-term state transitions. Which learning formulation most precisely matches this setting?

Types of machine learning techniques Hard
A. Unsupervised ranking with implicit clusters
B. Full-information online supervised learning
C. Contextual bandit learning with partial feedback
D. Model-based reinforcement learning with delayed states

47 A semi-supervised classifier assumes that nearby points usually share labels. Adding many unlabeled examples reduces its test accuracy. Which data geometry most directly explains this failure?

Types of machine learning techniques Hard
A. Labeled examples uniformly cover every cluster
B. Class boundaries lie only in low-density regions
C. Class boundaries pass through high-density regions
D. Unlabeled examples match labeled class proportions

48 In pool-based active learning, which condition most seriously undermines uncertainty sampling even when the classifier's probability estimates are calibrated on its current labeled set?

Types of machine learning techniques Hard
A. The labeled set contains examples from every class
B. The unlabeled pool contains out-of-distribution regions
C. The labeling oracle is costly but consistently accurate
D. The hypothesis class contains several nonlinear models

49 A hospital model predicts readmission using a feature named discharge_summary_length. At deployment, predictions must be produced upon admission, before any discharge summary exists. What is the primary defect?

Learning problems and system Hard
A. The model contains target leakage through unavailable future information
B. The model contains covariate shift through changing patient demographics
C. The model contains underfitting through insufficient feature complexity
D. The model contains class imbalance through rare readmission labels

50 Training and deployment satisfy but . If support overlap holds, which correction directly targets deployment risk?

Learning problems and system Hard
A. Weight each training loss by
B. Weight each training loss by
C. Weight each training loss by
D. Weight each training loss by

51 A fraud model is retrained weekly from labels that arrive 60 days after each transaction. A sudden fraud-strategy change occurs today. Which system limitation most directly delays adaptation?

Learning problems and system Hard
A. Batch inference prevents individual probability estimates
B. Model calibration preserves the previous decision threshold
C. Feature normalization uses historical summary statistics
D. Label latency postpones observable supervised feedback

52 A classifier retains the same ranking of all examples after deployment, but its predicted probabilities become systematically too large. Which property can remain unchanged while calibration degrades?

Learning problems and system Hard
A. Negative log-likelihood
B. Brier score
C. Area under the ROC curve
D. Expected calibration error

53 A preprocessing pipeline imputes missing values and standardizes features using the complete dataset before cross-validation. Why is the resulting cross-validation estimate optimistically biased?

Designing a learning system Hard
A. Each fold's model is optimized using a different loss function
B. Each fold's labels are stratified using the global class proportion
C. Each fold's training process uses statistics from its validation fold
D. Each fold's validation process uses fewer features than its training fold

54 A team compares 100 hyperparameter configurations by five-fold cross-validation and reports the best configuration's mean score as final performance. Which procedure best reduces selection bias while still tuning the model?

Designing a learning system Hard
A. Nested cross-validation with tuning restricted to inner folds
B. Bootstrap resampling of the original best fold predictions
C. Leave-one-out validation after selecting the highest score
D. Repeated training on all folds with the chosen parameters

55 A calibrated fraud model outputs . Investigating a transaction costs , while missing fraud costs ; a successful investigation prevents the entire fraud loss. When should the system investigate?

Designing a learning system Hard
A. Investigate when
B. Investigate when
C. Investigate when
D. Investigate when

56 Users who receive a model-triggered intervention become less likely to exhibit the measured adverse outcome. Training future models directly on these observed outcomes can create which system-level problem?

Designing a learning system Hard
A. Random label noise independent of model predictions
B. Intervention-induced feedback that changes observed labels
C. Static class imbalance unrelated to deployed decisions
D. Pure covariate shift with unchanged conditional outcomes

57 Using the formulation in which a program learns from experience with respect to tasks and performance measure , which specification is invalid because it confuses the task with the measure?

Concept of learning task Hard
A. : classify emails; : test-set ; : labeled emails
B. : minimize RMSE; : predict prices; : sale records
C. : choose actions; : return; : interaction trajectories
D. : rank documents; : NDCG; : relevance judgments

58 In noise-free concept learning, the version space is the set of hypotheses consistent with every observed example. After adding one correctly labeled example, which statement must hold?

Concept of learning task Hard
A. The version space must lose exactly one hypothesis
B. The version space must retain only the target concept
C. The version space can only shrink or remain unchanged
D. The version space can expand if the example is redundant

59 A concept learner searches only conjunctions of Boolean literals, but the true target is XOR of two attributes. With unlimited noise-free data, what prevents the learner from reaching zero population error?

Concept of learning task Hard
A. The labels violate stationarity in the target distribution
B. The examples violate independence in the sampling process
C. The learner violates calibration under zero-one loss
D. The target violates realizability in the hypothesis class

60 A finite hypothesis class is evaluated in a realizable PAC setting using a consistent learner. Which sample-size order is sufficient to achieve error at most with probability at least ?

Concept of learning task Hard
A.
B.
C.
D.