Guide-linked quiz · Medium Level
Empirical Risk Minimization Quiz
Covers empirical versus true risk, sampling assumptions, overfitting, regularized objectives, and why held-out evaluation matters.
Question 1 of 8
What does empirical risk minimization actually minimize during training?
Preview every question8 questions
Answers and explanations appear as you take the quiz.
- What does empirical risk minimization actually minimize during training?
- The true risk over the entire data distribution
- The number of parameters in the model
- The average loss computed over a finite training sample
- The variance of the model's predictions
- Why can't a model directly minimize true risk instead of empirical risk?
- True risk is defined over the entire, usually infinite and unknown data distribution, which is inaccessible
- True risk is always higher so it's not worth minimizing
- True risk cannot be expressed mathematically
- Models are only capable of computing averages
- As an IID sample grows, what classical result helps explain why its average loss can estimate expected loss for a fixed model?
- Bayes' theorem
- The central limit theorem alone
- Newton's method
- The law of large numbers along with uniform convergence bounds from VC theory
- During training, what does overfitting look like when empirical and true risk are compared?
- When both empirical and true risk are equally high
- When empirical risk is driven low, even to zero, while true risk remains high
- When a model refuses to train
- When the hypothesis class is too small to fit any data
- What role does regularization play in the ERM objective?
- It removes the loss function entirely
- It guarantees zero true risk
- It adds a penalty term that trades slightly higher empirical risk for a hypothesis less prone to overfitting
- It only affects the learning rate
- Which of the following is explicitly given as an implementation of the general ERM principle?
- Only neural networks
- Only linear regression
- None of these; ERM is a distinct algorithm separate from these
- Linear regression, logistic regression, SVMs, and neural network training, each with different loss functions
- What does structural risk minimization add beyond basic regularized ERM?
- It explicitly balances empirical risk against a measure of hypothesis-class complexity rather than fixing the class size
- It removes the need for a loss function
- It guarantees the model will never overfit
- It only applies to linear models
- In the ERM formula averaging L(h(x_i), y_i) over n examples, what does n represent?
- The number of features in the model
- The number of training examples
- The number of hypothesis classes considered
- The learning rate
Keep testing yourself
More quizzes picked for your level.