Statistical Learning Basic Concepts Practice Quiz (18 questions)

All questions and answers

1. A software system improves its performance by learning patterns from examples rather than receiving explicit instructions for each task. Which field studies this approach?

Statistical inference
Machine learning
Database management
Explicit rule-based programming

Machine learning

Explanation

Machine learning studies systems trained from data rather than explicitly programmed. Explicit rule-based programming relies on instructions supplied by a programmer, while statistical inference and database management are not the approach defined here.

2. A researcher has input-output examples and wants a model that predicts an output for a previously unseen input. What is the goal of this supervised-learning task?

Learn an input-output relation for predicting new outputs
Reconstruct the inputs in the training examples
Assign outputs independently of the input values
Estimate the entire joint distribution of inputs and outputs

Learn an input-output relation for predicting new outputs

Explanation

Supervised learning seeks an underlying relation between inputs and outputs that can predict an output for a new input. Reconstructing inputs, estimating the full joint distribution, or ignoring inputs does not express that goal.

3. A project has three prediction tasks: estimating a real-valued quantity, assigning one of two labels coded as −1-1 and 11, and choosing among TT categories labeled 11 through TT. Which output spaces match these tasks, respectively?

A subset of R\mathbb{R}; {−1,1}\{-1,1\}; {1,2,…,T}\{1,2,\ldots,T\}
A subset of R\mathbb{R}; {1,2,…,T}\{1,2,\ldots,T\}; {−1,1}\{-1,1\}
{−1,1}\{-1,1\}; a subset of R\mathbb{R}; {1,2,…,T}\{1,2,\ldots,T\}
{1,2,…,T}\{1,2,\ldots,T\}; {−1,1}\{-1,1\}; a subset of R\mathbb{R}

A subset of $$\mathbb{R}$$; $$\{-1,1\}$$; $$\{1,2,\ldots,T\}$$

Explanation

Regression uses real-valued outputs, binary classification uses the two labels −1-1 and 11, and multiclass classification uses labels from 11 through TT. The other combinations interchange the output spaces of these task types.

4. A team collects observations under a model in which every example is generated independently by the same fixed, unknown probability distribution. How should the sampling assumption be described?

Identically distributed sampling from a distribution known in advance
Dependent sampling from a fixed unknown distribution
Independent and identically distributed sampling from a fixed unknown distribution
Independent sampling from a distribution that changes with each observation

Independent and identically distributed sampling from a fixed unknown distribution

Explanation

The assumption is that the data are independently and identically sampled from one fixed but unknown distribution. A changing or known distribution, or dependence among observations, alters a key part of that assumption.

5. For a fixed input value xx, a model uses p(y∣x)p(y\mid x) to describe possible outputs and pX(x)p_X(x) to describe how likely that input is to be sampled. Which expression combines these into the joint distribution?

p(x,y)=pX(x)p(y∣x)p(x,y)=p_X(x)p(y\mid x)
p(x,y)=pX(x)+p(y∣x)p(x,y)=p_X(x)+p(y\mid x)
p(x,y)=p(y∣x)p(x,y)=p(y\mid x)
p(x,y)=pY(y)p(x∣y)p(x,y)=p_Y(y)p(x\mid y)

$$p(x,y)=p_X(x)p(y\mid x)$$

Explanation

The joint distribution factorizes into the input marginal multiplied by the conditional output distribution, p(x,y)=pX(x)p(y∣x)p(x,y)=p_X(x)p(y\mid x). Omitting the marginal, reversing the factorization, or adding the terms does not give the specified form.

6. A dataset has inputs sampled unevenly across the input space, and for a fixed input the output can still vary. Which distributions represent these two sources of uncertainty, respectively?

p(x,y)p(x,y) for input sampling and pX(x)p_X(x) for output variability
pX(x)p_X(x) for input sampling and p(y∣x)p(y\mid x) for output variability
p(y∣x)p(y\mid x) for input sampling and pX(x)p_X(x) for output variability
pX(x)p_X(x) for input sampling and p(x,y)p(x,y) for output variability

$$p_X(x)$$ for input sampling and $$p(y\mid x)$$ for output variability

Explanation

The marginal pX(x)p_X(x) describes uncertainty in which inputs are sampled, while p(y∣x)p(y\mid x) describes output uncertainty at a given input. Reversing the roles confuses the marginal with the conditional; using the joint distribution for just one role does not make the stated distinction.

7. A regression model uses a fixed but unknown function f∗f^* and adds random measurement noise ε\varepsilon to its output. Which expression represents the modeled response at input xx?

y=ε(x)+f∗y=\varepsilon(x)+f^*
y=f∗(x)+εy=f^*(x)+\varepsilon
y=f∗(x)y=f^*(x)
y=f∗(x)εy=f^*(x)\varepsilon

$$y=f^*(x)+\varepsilon$$

Explanation

The regression data model adds random noise to the value of the fixed, unknown regression function: y=f∗(x)+εy=f^*(x)+\varepsilon. Omitting noise removes the modeled variability, while treating the function as noise or multiplying by noise changes the model.

8. In a prediction task, what does a loss function assign to an actual output and a predicted output?

A prediction rule mapping inputs to outputs
A probability distribution over inputs
An average error over a data distribution
A nonnegative point-wise error

A nonnegative point-wise error

Explanation

A loss function maps an actual and predicted output to a nonnegative point-wise error. Input-sampling probability describes a distribution, a predictor maps inputs to outputs, and average error over a distribution is expected risk.

9. A predictor's loss is evaluated over a data distribution with joint density p(x,y)p(x,y). Which quantity gives its expected risk?

∫pX(x)ℓ(y,f(x)) dx\int p_X(x)\ell(y,f(x))\,dx
∫p(x,y)ℓ(y,f(x)) dx dy\int p(x,y)\ell(y,f(x))\,dx\,dy
∫p(x,y)f(x) dx dy\int p(x,y)f(x)\,dx\,dy
∫p(y∣x)ℓ(y,f(x)) dy\int p(y\mid x)\ell(y,f(x))\,dy

$$\int p(x,y)\ell(y,f(x))\,dx\,dy$$

Explanation

Expected risk averages the loss over the joint distribution of inputs and outputs: E(f)=∫p(x,y)ℓ(y,f(x)) dx dyE(f)=\int p(x,y)\ell(y,f(x))\,dx\,dy. Using only the input marginal or conditional omits part of the joint weighting, while averaging predictions does not calculate loss.

10. Two predictors are evaluated using the same loss and data distribution. One has the smallest expected risk among the available input-output functions; what role does it play?

It is the target function
It is the data distribution
It is the loss function
It is the training set

It is the target function

Explanation

For a given loss and data distribution, the target function is the input-output function that minimizes expected risk. A training-set estimator is produced from observed data, training error is not the defining criterion here, and a distribution describes the data rather than predicting outputs.

11. Why must a learning method estimate the target function from data rather than compute it directly?

The estimator is known before the training data are collected.
The distribution needed to determine the target function is unknown.
The target function is defined only for training examples.
The loss function cannot be specified before training.

The distribution needed to determine the target function is unknown.

Explanation

The target function minimizes expected risk under the data distribution, and that distribution is unknown, so the function generally cannot be computed directly. Treating the target as training examples, the loss as unknown, or the estimator as pre-given confuses distinct parts of the learning setup.

12. A procedure receives a training set and outputs a predictor tailored to that set. What is this procedure in the learning framework?

An estimator mapping the set to a target function
A target function mapping the set to a distribution
A data distribution mapping the set to an estimator
A learning algorithm mapping the set to an estimator

A learning algorithm mapping the set to an estimator

Explanation

A learning algorithm is the procedure that maps a training set to an estimator. Calling the output itself the algorithm confuses procedure with result; the target function and data distribution are not procedures that map training sets to estimators.

13. A predictor performs well on observed examples but differs substantially from the target function and has high excess expected risk. Which conclusion best fits the definition of generalization?

It generalizes well because it fits the observed examples.
It does not generalize well because its excess expected risk is not small.
It generalizes well because it was produced from a training set.
It does not generalize because its expected risk is not random.

It does not generalize well because its excess expected risk is not small.

Explanation

Generalization means that an estimator mimics the target function, formally corresponding to small excess expected risk. Good fit to observed data alone does not establish generalization, and being produced from a training set is not sufficient.

14. For a learning algorithm whose output depends on a random training set, which quantity measures the expected excess risk across possible training sets?

ES[E(f∗)−E(fS)]\mathbb{E}_S[E(f^*)-E(f_S)]
ES[E(f∗)]\mathbb{E}_S[E(f^*)]
E(fS)−E(f∗)E(f_S)-E(f^*) for one realized set
ES[E(fS)−E(f∗)]\mathbb{E}_S[E(f_S)-E(f^*)]

$$\mathbb{E}_S[E(f_S)-E(f^*)]$$

Explanation

The expected excess risk across training sets is the expectation over SS of the difference between the estimator's risk and the target function's risk. A single-set difference omits that averaging, while averaging target risk alone omits the excess and reversing the subtraction changes its sign.

15. As the sample size grows without bound, an algorithm's expected excess risk approaches zero. What property does this describe?

Sample complexity
A learning rate
Training-set fit
Consistency

Consistency

Explanation

Consistency is the property that expected excess risk tends to zero as the sample size tends to infinity. Sample complexity instead specifies a finite sample-size threshold for a chosen accuracy; a fixed risk level or exact equality at every finite size is not the consistency condition.

16. A learning algorithm fits its training data well, but small changes to the training examples can substantially change its estimator. Which design goal is missing for good learning?

A greater willingness to disregard the data
Stability under small changes to the data
A reduction in the number of training examples
A closer fit to the training data

Stability under small changes to the data

Explanation

Good learning requires both fitting the data well and remaining stable when the data change slightly. Treating fit as sufficient overlooks stability; confusing stability with fit conflates distinct goals; disregarding data describes oversmoothing, not the missing goal.

17. An estimator closely matches a training set, yet changes sharply when one example is slightly altered. How is the algorithm characterized?

Overfitting
Oversmoothing
Stable generalization
Weak data fitting

Overfitting

Explanation

Overfitting occurs when an estimator fits the data while being unstable. Oversmoothing instead combines stability with disregard for the data; stable generalization lacks the stated instability; weak data fitting contradicts the close match to the training set.

18. A researcher adjusts a parameter in a learning algorithm to balance how closely its estimator fits the data against how stable it remains. What role does this parameter play?

Controls the trade-off between data fitting and stability
Sets the number of training examples used
Controls stability without affecting data fitting
Controls the degree of data fitting without affecting stability

Controls the trade-off between data fitting and stability

Explanation

Regularization algorithms use one or more parameters to control the trade-off between data fitting and stability. The fit-only and stability-only choices omit the trade-off, while the sample-count choice confuses regularization with training-set size.

Study with flashcards

Lock in the answers with 35 flashcards on Statistical Learning Basic Concepts.

What does machine learning study?

Systems trained from data rather than explicitly programmed.

What relation does supervised learning seek to predict outputs for new inputs?

An underlying input-output relation.

In what space do inputs belong, and how is it bounded?

X⊆RDX \subseteq \mathbb{R}^D.

See the flashcards →

Read the study guide

Go over the full study guide for Statistical Learning Basic Concepts.

Read the study guide →

Related classes

Make your own practice quizzes

Upload your class notes and Revizly writes a practice quiz with explained answers in about a minute.

Quiz maker