Take the Deep Learning — Lecture 1: Introduction to ML and DL, Set 02: 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. In deep learning, what does the word "deep" refer to?
- A. Many successive layers of representation ✓
- B. The depth of mathematical theory behind the method
- C. The size of the training dataset
- D. How long the model takes to train
Answer: A. Many successive layers of representation
Explanation: "Deep" refers to many successive layers of representation, not to any deeper understanding.
Q2. How is the depth of a model defined?
- A. The number of training examples it has seen
- B. The number of layers that contribute to the model ✓
- C. The number of weights in its largest layer
- D. The number of classes it can predict
Answer: B. The number of layers that contribute to the model
Explanation: Depth is how many layers contribute to the model; modern networks use tens or hundreds.
Q3. What is shallow learning?
- A. Learning that uses only a small training set
- B. Learning that stops before the loss converges
- C. Machine-learning methods that learn only one or two layers of representation ✓
- D. Learning that uses no loss function
Answer: C. Machine-learning methods that learn only one or two layers of representation
Explanation: Other machine-learning methods learn one or two layers of representation, which is called shallow learning.
Q4. A deep network is described as an information-distillation operation. What does that mean?
- A. It compresses the training data to save memory
- B. It removes duplicate examples from the dataset
- C. It reduces the number of weights at every layer
- D. Every layer filters the information and keeps what is useful for the task ✓
Answer: D. Every layer filters the information and keeps what is useful for the task
Explanation: Information goes through successive filters and comes out increasingly purified for the task; every layer keeps what is useful.
Q5. In the cat-image example, what is the order of representations learned by successive layers?
- A. Objects, then textures and parts, then edges
- B. Raw pixels, then edges, then textures and parts, then objects ✓
- C. Raw pixels, then objects, then edges, then textures
- D. Edges, then raw pixels, then the final label
Answer: B. Raw pixels, then edges, then textures and parts, then objects
Explanation: The pipeline runs raw pixels, then edges, then textures and parts, then objects, giving the output "cat": 0.97.
Take the timed test (free, no sign-up) to see your score and the answer and explanation for every question.
Q6. What a layer does to its input is stored in its:
- A. Activation function
- B. Loss score
- C. Weights, also called parameters
- D. Training dataset
Q7. In this framing, what does learning mean?
- A. Choosing how many layers the network should have
- B. Collecting more labelled training data
- C. Selecting the right loss function
- D. Finding values for the weights of every layer
Q8. What does the loss function, also called the objective function, do?
- A. It compares the network's prediction with the true target and returns a distance score
- B. It chooses which layer to update next
- C. It decides how many epochs to train for
- D. It converts the input image into a tensor
Q9. What does the lecture call the fundamental trick in deep learning?
- A. Adding more layers until accuracy stops improving
- B. Using the loss score as a feedback signal to adjust the weights in the direction that lowers the score
- C. Initialising all weights to zero so training starts cleanly
- D. Training on the test set to maximise accuracy
Q10. Which component performs the weight adjustment, and which algorithm does it implement?
- A. The loss function, implementing gradient clipping
- B. The activation function, implementing normalisation
- C. The layer, implementing feature extraction
- D. The optimizer, implementing backpropagation
Q11. Why are the network's first predictions described as nonsense?
- A. The weights start at random values
- B. The loss function has not been chosen yet
- C. The training data has not been labelled yet
- D. The layers are not yet connected to each other
Q12. Which of these is listed under what deep learning has already achieved?
- A. Human-level natural-language understanding
- B. Superhuman Go playing
- C. Human-level general intelligence
- D. Believable dialogue systems
Q13. How does the lecture describe human-level general intelligence?
- A. Achieved in narrow domains only
- B. Expected within two years
- C. Still talk, not results
- D. Already solved by large language models
Q14. What hardware progression does the lecture give for deep learning, and what is claimed for the newest step?
- A. CPU to GPU to TPU, with TPUs reportedly 10 times faster and far more energy-efficient than top GPUs
- B. GPU to CPU to FPGA, with FPGAs being cheapest
- C. CPU to TPU to GPU, with GPUs being the final step
- D. Mainframe to cluster to cloud, with cloud being the cheapest
Q15. How is ImageNet described?
- A. 70,000 greyscale images of handwritten digits
- B. 1.4 million hand-annotated images in 1,000 categories, with a yearly competition
- C. 100 million molecules screened for antibiotic activity
- D. 200 million predicted protein structures
Q16. Which of these is named as a source of public medical datasets?
- A. ImageNet and MNIST
- B. Colab and Kaggle Notebooks
- C. TCIA and TCGA
- D. GitHub and Hugging Face Spaces
Q17. Which optimisers are named as part of the algorithmic advances that made deep learning work?
- A. Newton's method and simplex
- B. Genetic algorithms and simulated annealing
- C. Gradient boosting and random forests
- D. RMSProp and Adam
Q18. Which three advances are credited with allowing models thousands of layers deep to be trained from scratch?
- A. Batch normalisation, residual connections and depthwise separable convolutions
- B. Dropout, early stopping and data augmentation
- C. Layer-wise pre-training, autoencoders and restricted Boltzmann machines
- D. Transfer learning, fine-tuning and distillation
Q19. Under "the method itself is simpler to use", what does simplicity refer to?
- A. It needs fewer training examples than classical methods
- B. It removes the need for feature engineering
- C. It requires no loss function
- D. It can be written in fewer lines of code than any other method
Q20. What does versatility and reusability make possible, according to the lecture?
- A. Training without any labelled data
- B. Running the same model on any hardware without changes
- C. Guaranteeing the model will not overfit
- D. Continuous online learning, because a trained model can keep learning from new data without restarting