Quiz: The Foundations

Testing yourself is one of the strongest ways to remember anything. It's far better than rereading. So here's a challenge. Ten questions on the foundations of machine learning. Pause on each one, and really decide before you hear the answer. Keep your score, and let's see how you do.

• Testing beats rereading
• Ten questions
• Keep your own score

Question 1

Question 1. You want to predict tomorrow's temperature. What type of problem is this? A, Classification. B, Regression. C, Clustering. D, Reinforcement learning. Pause the video and choose your answer. The answer is B. A temperature is a number, so this is regression.

Question 2

Question 2. What does gradient descent try to reduce? A, The number of features. B, The loss. C, The learning rate. D, The size of the data. Pause the video and choose your answer. The answer is B. Every training step moves the weights so that the loss falls.

Question 3

Question 3. Your training score is ninety nine percent, and your test score is seventy. What's going on? A, Underfitting. B, Overfitting. C, Perfect fit. D, Data is too clean. Pause the video and choose your answer. The answer is B. A large gap between train and test means that the model has memorised the training data.

Question 4

Question 4. Which set must never be used to make modelling decisions? A, Training set. B, Cross validation folds. C, Test set. D, Features. Pause the video and choose your answer. The answer is C. The test set stays sealed until the very end, or its score stops being honest.

Question 5

Question 5. What's the main reason to use a pipeline? A, It looks tidy. B, It prevents leakage. C, It makes models bigger. D, It removes the need for data. Pause the video and choose your answer. The answer is B. Each preprocessing step is fitted only on the training folds, so nothing leaks.

Question 6

Question 6. Which model does not need feature scaling? A, K nearest neighbours. B, Support vector machine. C, Random forest. D, Logistic regression. Pause the video and choose your answer. The answer is C. Trees only ask whether a value is above or below a split, so scale doesn't matter.

Question 7

Question 7. Ninety nine percent of emails aren't fraud. A model gets ninety nine percent accuracy. What do you conclude? A, It is excellent. B, It may be doing nothing useful. C, It must be overfitting. D, It needs fewer features. Pause the video and choose your answer. The answer is B. Always predicting not fraud scores the same, so check recall and precision.

Question 8

Question 8. Lasso differs from ridge because it can do what? A, Make weights exactly zero. B, Train faster on images. C, Handle text. D, Remove the need for scaling. Pause the video and choose your answer. The answer is A. Lasso can switch features off completely, by setting their weights to zero.

Question 9

Question 9. Why do we shuffle before making cross validation folds? A, To add noise. B, The data may be ordered. C, It speeds up training. D, It removes outliers. Pause the video and choose your answer. The answer is B. Our housing data was in geographic order, which gave unfair folds. Time series are the exception.

Question 10

Question 10. Which is an example of target leakage? A, Using age to predict income. B, Using a treatment given after diagnosis to predict the disease. C, Scaling inside a pipeline. D, Using cross validation. Pause the video and choose your answer. The answer is B. That feature only exists because of the answer, so it won't be there at prediction time.

Summary