Skip to content

Quiz 4: Learning

CS50's Introduction to Artificial Intelligence with Python

Question 1

A social network's AI uses existing tagged photos to identify people in new photos. What type of machine learning is this?

  • This is an example of supervised learning ✓
  • This is an example of reinforcement learning
  • This is an example of unsupervised learning
  • This is not an example of machine learning

Why: The AI is trained on labeled input-output pairs (photos tagged with names). Learning from labeled examples to classify new inputs is the definition of supervised learning.

Question 2

Calculate the total L2 loss for five data points with the given true and predicted values.

Answer: 16 ✓

Why: L2 loss = Σ(actual − predicted)². Square each individual error and sum. Unlike L1 loss (absolute values), L2 penalizes larger errors more heavily.

Question 3

If Hypothesis 1 has lower L1 and L2 loss than Hypothesis 2 on training data, why might Hypothesis 2 still be preferable?

  • Hypothesis 1 might result from regularization
  • Hypothesis 1 might result from overfitting ✓
  • Hypothesis 1 might result from loss
  • Hypothesis 1 might result from cross-validation
  • Hypothesis 1 might result from regression

Why: A model with lower training loss isn't necessarily better — it may have memorized the training data (overfitting) and will generalize poorly to new examples. Hypothesis 2's higher training loss might mean it's a simpler, more generalizable model.

Question 4

In the ε-greedy approach to reinforcement learning, which ε value makes it identical to a purely greedy approach?

  • ε = 0 ✓
  • ε = 0.25
  • ε = 0.5
  • ε = 0.75
  • ε = 1

Why: With ε = 0, the probability of random exploration is 0% — the agent always exploits its best known action. With ε = 1, it would always explore randomly. The whole point of ε-greedy is balancing these extremes.