← Back to Week 3

Week 3: Building a Real-World Housing Price Predictor

Knowledge Check — Select an answer to see immediate feedback.

Questions 10

Question 1

When building a regression model for housing prices, the data pipeline typically includes a train/validation/test split. What is the purpose of the validation set, as distinct from the test set?

A) The validation set is used to train the model and the test set is used to tune hyperparameters
B) The validation set is used during development to tune hyperparameters and detect overfitting; the test set is held out and used only for final performance evaluation
C) The validation set replaces the test set in production deployments
D) Both the validation and test sets are used interchangeably to monitor training loss

Question 2

Feature scaling (e.g., standardization) is an important preprocessing step for neural networks. Why does failing to scale input features often harm training?

A) Unscaled features cause the loss function to become non-differentiable
B) Features with very different magnitudes cause the loss surface to be elongated, leading to inefficient gradient descent with slow convergence or oscillation
C) Unscaled features prevent the model from using all input features equally
D) Feature scaling is only necessary for convolutional networks, not fully connected networks

Question 3

R-squared (R²) is a way to measure how well a regression model explains the variation in the data. If a housing price model has an R² of 0.85, what does that mean?

A) The model makes an average error of 15% on each prediction
B) The model explains 85% of the variation in house prices — 15% remains unexplained
C) The model correctly predicts the price to within 15% accuracy 85% of the time
D) The model achieves 85% classification accuracy on the test set

Question 4

Overfitting is detected by comparing performance on the training and validation sets. Which pattern most clearly signals that a model is overfitting?

A) Training loss and validation loss both decrease steadily throughout training
B) Training loss decreases but validation loss plateaus or begins to increase after some point
C) Training loss increases while validation loss decreases
D) Both training and validation losses are high and do not decrease

Question 5

L2 regularization adds an extra penalty to the training loss that discourages the model from using very large weights. What is the practical effect of this?

A) It forces all weights to become exactly zero, shutting down unused neurons
B) It encourages the model to keep weights small, which tends to reduce overfitting and improve generalization
C) It increases the learning rate proportionally to how many weights the model has
D) It prevents the use of ReLU as an activation function

Question 6

Cosine annealing adjusts the learning rate according to a smooth decreasing schedule. What is the primary motivation for reducing the learning rate during training?

A) A decreasing learning rate prevents the model from using too much GPU memory
B) As training progresses, finer adjustments are needed to settle into a minimum without overshooting; a decaying rate allows coarse exploration early and fine refinement later
C) Cosine annealing ensures the learning rate never reaches zero, preventing the model from stopping
D) A decreasing learning rate automatically increases batch size to compensate

Question 7

In k-fold cross-validation, the dataset is split into k equally sized folds. The model is trained k times, each time using a different fold as the validation set. What is the primary benefit of this approach over a single train/validation split?

A) K-fold cross-validation trains k separate models that are then combined for better predictions
B) It provides a more reliable performance estimate with lower variance, since every data point is used for validation exactly once across the k iterations
C) It eliminates the need for a separate test set
D) It automatically prevents overfitting by duplicating the training data

Question 8

For a regression task like predicting housing prices, what should the output layer of a neural network look like?

A) A single neuron with a sigmoid activation function outputting a value between 0 and 1
B) Multiple neurons with softmax activation, one per price category
C) A single neuron with no activation function (linear output), allowing the model to predict any continuous real value
D) A single neuron with ReLU activation to ensure non-negative price predictions

Question 9

MSE and MAE are both regression metrics, but they respond differently to outliers. Which statement best captures this difference?

A) MSE is less sensitive to outliers because it averages the errors rather than squaring them
B) MAE treats all prediction errors equally regardless of magnitude, making it more robust to outliers than MSE, which amplifies large errors quadratically
C) Both MSE and MAE are equally sensitive to outliers since they both average across all samples
D) MSE is preferred when the dataset has outliers because it down-weights large errors

Question 10

Error analysis involves examining cases where a model makes the largest mistakes. What is the most useful outcome of conducting a thorough error analysis?

A) It provides a way to remove outliers from the dataset automatically
B) It reveals patterns in the types of examples the model struggles with, guiding targeted improvements such as collecting more data, engineering better features, or adjusting the architecture
C) It proves that the model has converged to a global minimum of the loss function
D) It automatically generates new training data to fill gaps in model knowledge