Ensemble Learning

Question 1

Was ist Ensemble Learning?

Answer 1

select a collection (ensemble) of hypotheses (models) and combine their predictions

Question 2

Wie kann man Diversität für Ensemble Learning erreichen?

Answer 2

• Using different learning algorithms
• Using different hyper-parameters in the same algorithm
• Using different input representations, e.g. different
  subsets of input features
• Using different subsets of training data (bagging, boosting and cascading)

Question 3

Erkläre, wie diversity from differences in input features erreicht wird

Answer 3

Question 4

Erkläre, wie Diversity from subsets of training data erreicht wird

Answer 4

Question 5

Wie kann man den Output von Base Learner beim Ensemble Learning kombinieren?

Answer 5

Question 6

Wieso ist der ensemble error geringer als der individuelle error?

Answer 6

Question 7

Wieso hat man einen geringeren Error of Ensemble via Voting?

Answer 7

Question 8

Erkläre die Rank-Level Fusion Method

Answer 8

Question 9

Was sind die Vorteile von Bagging?

Answer 9

• For noisy data: not considerably worse, more robust
• Improved accuracy in prediction
• Decreases error by decreasing the variance in the results
  due to unstable learners

Question 10

Erkläre die Random Subspace Method

Answer 10

Question 11

Wieso ist die random subspace method sinnvoll für random forests?

Answer 11

this prevents different trees from choosing the same features for splits, which would make them highly correlated

Question 12

Wofür wird die Dropout Technik genutzt? Und was ist das?

Answer 12

Um zu vermeiden, dass NNs overfitten

Question 13

Wie funktioniert Boosting?

Answer 13

Question 14

Boosting

Was ist der Unterschied zwischen Strong und Weak Learners?

Answer 14

Question 15

Wie kann man weak learners boosten, sodass sie ein strong learner werden?

Answer 15

Question 16

Wie funktioniert AdaBoost (Adaptive Boosting)?

Answer 16

Question 17

Was ist der Vorteil von Boosting gegenüber von Bagging?

Answer 17

1. Needs smaller number of training samples than bagging
2. faster computation
3. higher accuracy

Question 18

Was sind Vorteile von Boosting?

Answer 18

Question 19

Was ist Gradient Boosting? Was ist der Unterschied zu AdaBoost?

Answer 19

Question 20

Welche Laufzeit hat Boosting für Face Detection?

Answer 20

Question 21

Beschreib Attentional Cascades

Answer 21

Chain classifiers that are progressively more complex and
have lower false positive rates

Question 22

Vergleiche AdaBoost mit MLPs

Answer 22

Question 23

Vergleiche Bagging mit Boosting.
Wie wird eine Entscheidung via Voting getroffen?

Answer 23

Question 24

AdaBoost can be used both to select a subset of informative features and to act as classifier for face detection problems.
Stimmt das?

Answer 24

Ja

Question 25

The information stored in a replay buffer depends on the learned parameters of the agent.
Stimmt das?

Answer 25

Nein

Question 26

In a markov decision process, the probability of moving from state s to state s’ by action a can depend on the actions leading to state s.
Stimmt das?

Answer 26

Nein

Question 27

Which of the following statements on RL algorithms are correct?
1. SARSA updates are based on the chosen action.
2. In Q-Learning, an episode has to be played to the end before it is possible to update the Q-values.
3. Q-learning is an on-policy algorithm.
4. SARSA uses a monte-carlo estimate of the return to calculate the Q-values.

Answer 27

1

Question 28

In Q-Learning, the TD error is the difference between the predicted quality of a state action pair 𝑄(𝑠,𝑎)
and the bootstrapped return which takes the actual reward for action a in state s into account.
Stimmt das?

Answer 28

Ja

Question 29

Bagging, Boosting or both?

Answer 29