Armin

Question 1

Bias-variance trade-off

Answer 1

The tradeoff between bias and variance. In many scenarios, adding bias can decrease the variance in a model. That result in a more accurate prediction. But there is always a tradeoff in findin a sweet spot that can be hard to find. This can however be addressed with cross-validation

Question 2

Consistency (or lack thereof) in shrinkage methods

Answer 2

If the number of parameters are higher than the number of observations. Shrinkage methods can be used such as shrinkage or lasso. With shrinkage methods we avoid a potential singular matrix we add a penalty term

Svarar ej på frågan?

Question 3

Support consistency of lasso

Answer 3

????

Question 4

Linear classifiers and perceptron

Answer 4

Classifiers that classify by trying to separate the groups by a straight line or hyperplane

Question 5

Generalization error

Answer 5

A measure on how accuracy an algorithm is able to predict outcomes of previously unseen data.

Question 6

Rademacher complexity

Answer 6

???

Question 7

Population risk

Answer 7

Expect errors for the loss-function given a known distribution. So to calculate the risk we need to know the distribution.

Question 8

Empirical Risk

Answer 8

The amount of wrongly classified observations for test-data when we don’t know the underlying distribution

Risk = Observations/ misclassifications