Take the Deep Learning — Lecture 1: Introduction to ML and DL, Set 03: 20 multiple-choice questions (30 minutes) with instant scoring and a full answer…
The first 5 questions of this test, with the correct answer and explanation.
Q1. What assumption gives Naive Bayes its name?
- A. It assumes the training data is free of noise
- B. It assumes every input feature is independent of the others ✓
- C. It assumes all classes are equally likely
- D. It assumes the data follows a normal distribution
Answer: B. It assumes every input feature is independent of the others
Explanation: Naive Bayes assumes every input feature is independent of the others, a strong ("naive") assumption, which is where the name comes from.
Q2. What does the lecture point out about logistic regression?
- A. It only works on image data
- B. It is a deep learning method
- C. It cannot be used as a baseline
- D. It is a classification algorithm despite its name ✓
Answer: D. It is a classification algorithm despite its name
Explanation: Logistic regression is a close relative of Naive Bayes and is a classification algorithm despite its name.
Q3. What is the practical tip given about classical probabilistic models?
- A. Try them first as a baseline; if a deep model cannot beat them, you have learned something about your data ✓
- B. Avoid them entirely, since deep learning always wins
- C. Use them only after the deep model has been fully tuned
- D. Report them instead of the deep model if they are faster
Answer: A. Try them first as a baseline; if a deep model cannot beat them, you have learned something about your data
Explanation: The tip is to try these first as a baseline; if a deep model cannot beat them, you have learned something about your data.
Q4. How does a support vector machine choose its decision boundary?
- A. By minimising the number of support vectors
- B. By picking the boundary closest to the mean of each class
- C. By choosing the hyperplane that maximises the distance to the closest points of each class ✓
- D. By randomly sampling boundaries and keeping the best
Answer: C. By choosing the hyperplane that maximises the distance to the closest points of each class
Explanation: An SVM chooses the hyperplane that maximises the distance to the closest points of each class, which is maximising the margin.
Q5. What is the first step in how an SVM works?
- A. Reduce the data to two dimensions for plotting
- B. Map the data to a new high-dimensional representation where the decision boundary is a hyperplane ✓
- C. Remove all outliers from the training set
- D. Cluster the data into groups before classifying
Answer: B. Map the data to a new high-dimensional representation where the decision boundary is a hyperplane
Explanation: The SVM maps the data to a new high-dimensional representation where the decision boundary becomes a hyperplane, a straight line in 2D.
Take the timed test (free, no sign-up) to see your score and the answer and explanation for every question.
Q6. Which of the following is NOT given as a reason kernel methods lost ground to deep learning?
- A. They are hard to scale to large datasets
- B. They are weak on perceptual problems such as image classification
- C. They need useful features extracted by hand
- D. They cannot produce a decision boundary at all
Q7. Which shallow methods does the lecture say are still excellent on tabular data?
- A. Naive Bayes, k-means and PCA
- B. Convolutional networks and recurrent networks
- C. Decision trees, random forests and gradient boosting machines
- D. Autoencoders and restricted Boltzmann machines
Q8. How many protein structures did AlphaFold predict and release free to biologists in 2022?
- A. 200 million
- B. 1.4 million
- C. 100 million
- D. 70,000
Q9. What was cleared by the US FDA in 2018 as the first autonomous AI diagnostic?
- A. Lung cancer detection from CT scans
- B. Skin cancer classification from phone photos
- C. Diabetic retinopathy detected from a retinal photo, with no specialist needed
- D. Heart disease prediction from ECG data
Q10. What is Halicin, and how was it found?
- A. A protein structure predicted by AlphaFold in 2022
- B. A weather model that forecasts 10 days ahead
- C. A dataset of 100 million annotated molecules
- D. An antibiotic found in 2020 by a neural network screening over 100 million molecules
Q11. What did GraphCast achieve in 2023?
- A. A 10-day global weather forecast in under a minute, beating the leading physics-based system on most measures
- B. The first AI-designed vaccine
- C. Real-time translation of 100 languages
- D. Superhuman performance at Go
Q12. What common thread does the lecture draw across AlphaFold, the FDA diagnostic, Halicin and GraphCast?
- A. All four were built by the same company
- B. A large dataset, a deep network, and a loss function
- C. All four required a supercomputer costing millions
- D. All four replaced human experts completely
Q13. Why does the lecture call local data an advantage for students in Pakistan?
- A. Local data is always cheaper to label
- B. Local data needs no ethical approval
- C. Local models train faster than international ones
- D. A dataset of Pakistani crops, dialects or hospitals does not exist yet, so the first good one gets cited for years
Q14. What is the four-step path from a public dataset to a published paper?
- A. Pick a dataset, run a baseline, improve one thing, write and submit
- B. Write the paper, find a dataset, run experiments, submit
- C. Collect data, buy compute, hire a team, publish
- D. Read the literature, propose a theory, prove it, submit
Q15. What does the lecture say about dataset papers?
- A. They are rarely accepted by journals
- B. They are often cited more than a model paper
- C. They cannot be published without a new model
- D. They require no experiments at all
Q16. Which of these is named as a source of free compute for researchers?
- A. HEC NRPU grants
- B. Ignite National Technology Fund
- C. Google TPU Research Cloud
- D. Industry data partnerships with hospitals
Q17. What pattern does the lecture state about earnings from machine-learning skills?
- A. Earnings follow the number of certificates you hold
- B. Earnings follow the university you graduated from
- C. Earnings depend mainly on which country you live in
- D. Earnings follow finished projects, so every assignment is a portfolio piece
Q18. Which tool is recommended for adding and removing layers and watching a network learn to separate spirals?
- A. TensorFlow Playground
- B. Teachable Machine
- C. Hugging Face Spaces
- D. Google Colab
Q19. In the quiz, an app that identifies a mango-leaf disease from a raw phone photo is classified as deep learning. Why?
- A. Because it runs on a phone rather than a server
- B. Raw pixels go in and a label comes out: the layers learn the features themselves
- C. Because the training set is very large
- D. Because a human chose the 30 most useful features
Q20. How is the course graded?
- A. 50% assignments and 50% exams
- B. 100% exams
- C. 20% attendance and homework, 80% exams
- D. 30% attendance, 30% homework, 40% exams