← Back to Week 1

Week 1: Introduction to Deep Learning

Knowledge Check — Select an answer to see immediate feedback.

Questions 10

Question 1

What is the primary advantage of deep learning over traditional machine learning approaches?

A) Deep learning always trains faster than classical algorithms
B) Deep learning automatically learns hierarchical feature representations without manual feature engineering
C) Deep learning requires less data than traditional ML methods
D) Deep learning models are always simpler and more interpretable

Question 2

During training, a neural network adjusts its weights step by step based on its mistakes. What controls how big each adjustment step is?

A) The number of training examples in the dataset
B) The learning rate — a setting that determines how large or small each update to the weights is
C) The number of layers in the network
D) The activation function used in the output layer

Question 3

Sigmoid is an activation function often used in the output layer for binary classification. What kind of output does it produce?

A) Any number from negative to positive (unbounded)
B) Only exactly 0 or 1 — a hard binary decision
C) A number between 0 and 1, which can be interpreted as a probability
D) A number between -1 and 1

Question 4

Which of the following best describes a key limitation of the sigmoid activation function that motivated the adoption of ReLU?

A) Sigmoid outputs are not bounded between 0 and 1
B) Sigmoid suffers from the vanishing gradient problem, slowing learning in deep networks
C) Sigmoid is computationally more expensive than any other activation function
D) Sigmoid cannot be used in the output layer of classification networks

Question 5

Which loss function is most appropriate for a binary classification task?

A) Mean Squared Error (MSE)
B) Binary Cross-Entropy
C) Categorical Cross-Entropy
D) Mean Absolute Error (MAE)

Question 6

PyTorch uses dynamic computation graphs, while early TensorFlow (before version 2.0) used static computation graphs. What is a key practical benefit of dynamic graphs?

A) Static graphs are always faster at inference time
B) Dynamic graphs allow the network structure to be defined and modified on-the-fly, making debugging more intuitive
C) Dynamic graphs require less GPU memory than static graphs
D) Dynamic graphs automatically tune the learning rate

Question 7

Three major categories of machine learning are supervised, unsupervised, and reinforcement learning. Which of the following tasks is an example of unsupervised learning?

A) Image classification using labeled training data
B) Predicting house prices from historical sales records
C) Discovering customer segments in purchase data without predefined labels
D) Training a game-playing agent using win/loss rewards

Question 8

When you first create a neural network, all the weights have to start at some initial values. Why does the choice of starting values actually matter?

A) It doesn't matter — the network will always learn the same thing regardless of where it starts
B) Poor starting values can cause signals in the network to blow up or shrink to nothing, making it impossible to train
C) The starting values permanently determine the final accuracy and cannot be improved by training
D) Weights must always start at exactly zero to ensure fairness across neurons

Question 9

Which framework has historically been preferred in production and large-scale deployment environments, and why?

A) PyTorch, because it was released first and has the largest ecosystem
B) TensorFlow, due to its static computation graph optimizations and tools like TensorFlow Serving and TensorFlow Lite
C) PyTorch, because static graphs are better suited for production
D) TensorFlow, because it cannot be used for research purposes

Question 10

Mean Squared Error (MSE) squares the difference between a prediction and the true value before averaging. What effect does squaring have on how the model treats errors?

A) It cancels out over-predictions and under-predictions so they balance each other
B) It ensures the loss value is always negative, rewarding correct predictions
C) It penalizes big errors much more heavily than small ones, making the model sensitive to outliers
D) It makes the calculation run faster on the GPU