Pre-test: Introduction to Machine Learning

Part 1

1) Who developed the Perceptron, one of the earliest neural network models?

Frank Rosenblatt Marvin Minsky Geoffrey Hinton Alan Turing

NoteExplanation

Frank Rosenblatt introduced the Perceptron (1957), an early neural network model, which inspired later deep learning research.

2) Who coined the term “Artificial Intelligence” at the Dartmouth Conference in 1956?

Alan Turing Herbert Simon John McCarthy Marvin Minsky

NoteExplanation

John McCarthy is credited with coining the term Artificial Intelligence and organizing the Dartmouth Conference in 1956.

3) Who created the Support Vector Machine (SVM) algorithm?

Vladimir Vapnik and Alexey Chervonenkis Geoffrey Hinton Yann LeCun Donald Knuth

NoteExplanation

Vladimir Vapnik and Alexey Chervonenkis introduced SVM in the 1990s, which became a powerful tool for classification.

4) Who is the founder of DeepMind, the team behind AlphaGo?

Demis Hassabis Geoffrey Hinton Ian Goodfellow Yoshua Bengio

NoteExplanation

Demis Hassabis co-founded DeepMind, which developed AlphaGo and AlphaZero using deep reinforcement learning.

5) Why is Arthur Samuel’s checkers program considered a milestone?

it was the first program that only followed hand-coded rules it was the first program that could learn to improve its own play it was the first program to apply probability in board games it was a program originally designed by Alan Turing

NoteExplanation

Arthur Samuel’s checkers program (1950s) was the first successful self-learning program, a landmark in machine learning history.

6) Which example shows applied machine learning?

training an algorithm to predict customer churn based on past records summarizing patterns from data without building a model studying the mathematical foundations of ML without application writing a report comparing algorithm advantages conceptually

NoteExplanation

Applied ML means building and training models from data. Theory alone is not applied practice.

7) How does machine learning relate to AI?

machine learning is one subfield under the larger umbrella of AI machine learning is broader and includes all of AI AI and ML are two separate, unrelated areas AI is a subset contained within machine learning

NoteExplanation

AI is the broad field; ML is a subset focused on learning from data.

8) What best characterizes supervised learning?

learning patterns without any target values learning exclusively through artificial intelligence hardware learning from labeled data where correct outputs are provided learning designed to mimic human creativity

NoteExplanation

Supervised learning uses input–output pairs (labeled data) to train models.

9) Which statement best defines machine learning?

a computer’s ability to memorize experiences a computer system that learns patterns from data a system that fully understands human thoughts a program that only replays past results without learning

NoteExplanation

ML is about improving from data automatically rather than coding every rule.

10) What is the key purpose of classification?

to gather raw data records to explore attributes without predicting to assign unseen examples into discrete categories to manually tag each data point

NoteExplanation

Classification models predict classes such as spam/ham or positive/negative.

11) What is the objective of unsupervised learning?

to train using labeled outcomes only to uncover hidden structures or patterns in unlabeled data to classify with pre-labeled examples to fine-tune hyperparameters of a supervised model

NoteExplanation

Unsupervised learning discovers structure (clusters, associations) without labels.

12) What is the goal of association rule learning?

to find natural groupings or clusters in data to discover rules about frequent co-occurring items in data to forecast outcomes using labeled training sets to classify new data into pre-existing categories

NoteExplanation

Association rule mining (e.g. market basket analysis) finds frequent item relationships like Milk → Bread.

13) Which algorithm repeatedly assigns data points into k clusters?

hierarchical clustering k-means clustering DBSCAN agglomerative clustering

NoteExplanation

K-means iteratively assigns points to centroids and updates until convergence.

14) Which is a typical use case of reinforcement learning?

training self-driving cars to make sequential driving decisions predicting tomorrow’s weather with regression grouping customers by demographics estimating fruit prices with linear regression

NoteExplanation

Reinforcement learning solves sequential decision-making problems with rewards, such as autonomous driving.

15) Which learning paradigm uses both labeled and unlabeled data?

supervised learning unsupervised learning semi-supervised learning reinforcement learning

NoteExplanation

Semi-supervised methods leverage a small labeled set plus a large unlabeled set to improve learning.

16) Which of the following is an evaluation metric for classification?

mean squared error (MSE) accuracy R-squared sum of squared errors (SSE)

NoteExplanation

Accuracy = (correct predictions) / (total predictions), a standard classification metric.

17) Which algorithm is a common baseline for linear classification?

k-means logistic regression apriori random forest regressor

NoteExplanation

Logistic regression models class probabilities for binary/multiclass settings.

18) Which technique is best for grouping customers by similar behavior?

regression classification clustering reinforcement learning

NoteExplanation

Clustering groups observations by similarity without labels (e.g., k-means, hierarchical).

19) What is overfitting in machine learning?

the model fits the training data poorly the model fits training data too closely and performs poorly on new data the model generalizes perfectly the model ignores training patterns

NoteExplanation

Overfitting = low bias, high variance; excellent on training data but weak generalization on unseen data.

20) Which technique helps reduce overfitting in machine learning?

using a larger and more complex model cross-validation ignoring validation data training for infinitely many epochs

NoteExplanation

Cross-validation evaluates the model on multiple data splits, helping detect and reduce overfitting.

21) Which of the following is an example of regression in ML?

classifying emails as spam or not spam predicting house prices grouping customers by similarity detecting fraudulent transactions

NoteExplanation

Regression predicts continuous values, such as house prices or temperature.

22) Which of these is an unsupervised learning algorithm?

decision trees k-means clustering linear regression naive Bayes

NoteExplanation

K-means is unsupervised because it finds groups in unlabeled data.

23) What does a confusion matrix show in classification?

only loss values training accuracy only true/false positives and negatives gradient descent weight updates

NoteExplanation

A confusion matrix summarizes results into TP, TN, FP, FN, giving more detail than accuracy alone.

24) Which optimization algorithm is most common in training neural networks?

apriori gradient descent random walk k-nearest neighbors

NoteExplanation

Gradient descent updates model weights iteratively to minimize loss.

25) Which ML model is inspired by the human brain?

logistic regression artificial neural networks random forest k-means clustering

NoteExplanation

Neural networks are modeled after biological neurons connected in layers.

26) Which task is NOT supervised learning?

sentiment analysis predicting exam scores clustering students by study style spam email classification

NoteExplanation

Clustering is unsupervised; the other tasks use labeled data, so they’re supervised.

27) What is the purpose of feature scaling?

to ensure features have similar ranges for training to increase dataset size to reduce memory usage to eliminate irrelevant features

NoteExplanation

Scaling prevents features with larger numeric ranges from dominating distance-based or gradient-based models.

28) Which distance metric is most common in KNN?

hamming distance euclidean distance mahalanobis distance cosine similarity

NoteExplanation

Euclidean distance is the standard metric in KNN to measure closeness.

29) What does PCA (Principal Component Analysis) do?

builds classification models reduces dimensionality by extracting key features clusters data points predicts labels for targets

NoteExplanation

PCA compresses data into fewer dimensions while keeping most variance.

30) Which is true about Random Forest?

uses a single decision tree combines multiple decision trees for better accuracy only works for regression requires no randomness

NoteExplanation

Random Forest is an ensemble of many decision trees that improves stability and accuracy.

31) Which task is most suitable for reinforcement learning?

training a robot to walk predicting house prices detecting spam emails customer segmentation

NoteExplanation

RL handles sequential decisions like a robot learning to walk through trial and error.

32) Which is a common loss function for regression models?

cross-entropy loss mean squared error (MSE) hinge loss log loss

NoteExplanation

MSE is the standard regression loss measuring average squared differences.

33) Which dataset split is used to test generalization?

training set validation set test set raw dataset

NoteExplanation

The test set is held back until the end to check how well the model generalizes to unseen data.

34) What is a limitation of K-means clustering?

always finds the global optimum requires specifying the number of clusters in advance cannot handle numerical data only works on small datasets

NoteExplanation

K-means needs k predefined, which is often unknown and requires experimentation.

Part 2

Q1: Predicting house prices from historical data can be solved using Supervised Learning. TRUEFALSE

NoteExplanation

House-price prediction maps features (e.g., size, location) to a numeric target (price). This is supervised regression because the model learns from labeled examples (x, y).

Q2: Which algorithm is commonly used for classification in Supervised Learning?

K-means Logistic Regression Q-learning Apriori

NoteExplanation

Logistic Regression is a classic supervised classification algorithm.

• K-means = unsupervised clustering
• Q-learning = reinforcement learning
• Apriori = association rule mining (unsupervised)

Q3: Sentiment analysis of customer reviews is an example of Supervised Learning. TRUEFALSE

NoteExplanation

Sentiment analysis uses labeled text (e.g., positive/negative/neutral). Labels make it supervised classification.

Q4: Which is required for Supervised Learning?

Labeled data Unlabeled data Rewards and penalties None of the above

NoteExplanation

Supervised learning needs labeled data (inputs with known outputs).

• Unlabeled → typically unsupervised
• Rewards/penalties → reinforcement learning

Q5: Supervised Learning can be applied to spam email detection. TRUEFALSE

NoteExplanation

Spam detection is binary classification trained on emails labeled spam/not spam.

Q6: Market segmentation using customer purchase data is an example of Unsupervised Learning. TRUEFALSE

NoteExplanation

Segmentation often uses clustering (e.g., K-means) to discover groups without labels → unsupervised.

Q7: Which algorithm belongs to Unsupervised Learning?

Linear Regression K-means Clustering Decision Tree Q-learning

NoteExplanation

K-means is an unsupervised clustering method. Linear Regression & Decision Tree (supervised), Q-learning (reinforcement).

Q8: In Unsupervised Learning, the algorithm is trained with both inputs and labeled outputs. TRUEFALSE

NoteExplanation

Unsupervised learning uses inputs only (no target labels) to find structure (clusters, components).

Q9: Which of the following is NOT an application of Unsupervised Learning?

Image compression Customer segmentation House price prediction Dimensionality reduction

NoteExplanation

House price prediction is supervised regression. Image compression & dimensionality reduction (e.g., PCA) and segmentation are unsupervised.

Q10: Association rule mining (e.g., “Customers who buy bread also buy butter”) is part of Unsupervised Learning. TRUEFALSE

NoteExplanation

Association rules (e.g., Apriori) learn co-occurrence patterns from unlabeled transaction data.

🔹 Reinforcement Learning

Q11: A robot learning to walk by trial and error is an example of Reinforcement Learning. TRUEFALSE

NoteExplanation

In RL, an agent interacts with an environment, receives rewards/penalties, and iteratively improves its policy.

Q12: In Reinforcement Learning, the agent learns by:

Labeled training data Grouping data points Receiving rewards or penalties Direct supervision

NoteExplanation

RL optimizes behavior via reward signals, not labels or grouping.

Q13: Reinforcement Learning is suitable for training self-driving cars. TRUEFALSE

NoteExplanation

Self-driving involves sequential decisions with delayed outcomes—well-suited to RL (often combined with supervised/perception models).

Q14: Which of the following best describes Reinforcement Learning?

Learning with labels Learning without labels Learning by interacting with the environment Learning by clustering data

NoteExplanation

RL = interaction + feedback (rewards). It’s neither standard labeled learning nor pure clustering.

Q15: Game-playing AI like AlphaGo uses Reinforcement Learning. TRUEFALSE

NoteExplanation

AlphaGo/AlphaZero use RL (self-play, policy/value networks) to maximize winning probability.

Q16: Which learning type would you use for fraud detection in banking?

Supervised Unsupervised Reinforcement

NoteExplanation

Fraud detection is commonly supervised classification with labeled fraud/non-fraud. Note: Unsupervised anomaly detection can complement when labels are scarce.

Q17: Which learning type would you use for grouping students by similar learning styles?

Supervised Unsupervised Reinforcement

NoteExplanation

Grouping without known labels → unsupervised clustering (e.g., K-means, hierarchical clustering).

Q18: Which learning type would you use for a robot navigating a maze?

Supervised Unsupervised Reinforcement

NoteExplanation

Maze navigation requires sequential decision-making with rewards (e.g., reaching the goal) → RL.

Q19: Predicting tomorrow’s temperature using past weather data is an application of Unsupervised Learning. TRUEFALSE

NoteExplanation

Forecasting a numeric target from historical features is supervised regression (often time-series models).

Q20: Teaching an AI to play chess without providing the rules but only giving rewards for wins is Reinforcement Learning. TRUEFALSE

NoteExplanation

The agent learns policies from reward signals (win/loss) through environment interaction, i.e., reinforcement learning.