Knowledge check
Knowledge check
Evaluate your understanding of this unit by completing the knowledge check below.

Action item: Knowledge check
Work through each question. Correct answers and feedback mirror the Multiverse assessment.
- A. To memorise all patterns in the training data for maximum accuracy.
- B. To identify and store all training examples for use in predictions.
- C. To adjust model parameters to minimize the loss function and improve predictions.
- D. To reduce computational costs by skipping gradient updates.
The best choice is (C): To adjust model parameters to minimize the loss function and improve predictions.
- A. The direction in which the gradient points.
- B. The number of iterations required to reach convergence.
- C. The size of the step taken toward the minimum of the loss function.
- D. The total number of features used in training.
The best choice is (C): The size of the step taken toward the minimum of the loss function.
- A. It updates weights too frequently for large datasets.
- B. It struggles with slow convergence and can get stuck in local minima.
- C. It automatically adjusts learning rates for each parameter.
- D. It is only effective for non-linear models.
The best choice is (B): It struggles with slow convergence and can get stuck in local minima.
- A. Adagrad
- B. SGD with Nesterov acceleration.
- C. Adadelta
- D. Adam (Adaptive Moment Estimation)
The best choice is (D): Adam (Adaptive Moment Estimation)
- A. A large, noisy dataset where gradients vary significantly between batches.
- B. A small, clean dataset with stable gradients.
- C. A dataset with sparse features and rare signals.
- D. A small dataset requiring strong feature selection.
The best choice is (A): A large, noisy dataset where gradients vary significantly between batches.
- A. To reduce the influence of correlated features without removing them.
- B. To eliminate irrelevant features by forcing their coefficients to zero.
- C. To maintain all features but shrink their coefficients.
- D. To reduce computation time during gradient updates.
The best choice is (B): To eliminate irrelevant features by forcing their coefficients to zero.
- A. The model converges faster to the global minimum.
- B. The model perfectly fits both training and test data.
- C. The model ignores the bias term.
- D. The model may underfit the data.
The best choice is (D): The model may underfit the data.
- A. It permanently removes neurons from the network to simplify architecture.
- B. It increases the learning rate for underperforming neurons.
- C. It randomly deactivates neurons during training to prevent over-reliance on specific pathways.
- D. It duplicates neurons to create redundant layers.
The best choice is (C): It randomly deactivates neurons during training to prevent over-reliance on specific pathways.
- A. L2 (Ridge)
- B. L1 (Lasso)
- C. Early stopping
- D. Dropout
The best choice is (A): L2 (Ridge)
- A. Underfitting — the model hasn’t learned enough from the training data.
- B. Data leakage — the test data was used during training.
- C. Overfitting — the model has learned noise from the training data.
- D. Poor optimiser selection — the model didn’t converge.
The best choice is (C): Overfitting — the model has learned noise from the training data.