Fundamentals

Question 1

What are parametric models?

Answer 1

Models with a fixed number of parameters

Question 2

What are non-parametric models?

Answer 2

Models for which the number of parameters is not fixed, it can increase when more data arrives

Question 3

What are examples of non-parametric models?

Answer 3

KNN, decision trees, random forests, kernel methods, Gaussian processes

Question 4

What are the three components of a supervised learning model

Answer 4

Model type, cost function, optimizer

Question 5

What is supervised learning?

Answer 5

A type of ML where the model is trained on labeled data to predict outcomes.

Question 6

What is unsupervised learning?

Answer 6

ML where the model finds hidden patterns in unlabeled data.

Question 7

Explain bias-variance trade-off.

Answer 7

Bias: error from overly simplistic model. Variance: error from sensitivity to training data. Trade-off is balancing both for best generalization.

Question 8

What is L1 regularization?

Answer 8

Adds the sum of absolute values of weights to the loss function; encourages sparsity.

Question 9

What is L2 regularization?

Answer 9

Adds the sum of squared weights to the loss function; discourages large weights.

Question 10

What is cross-validation?

Answer 10

A technique for assessing model performance by training/testing on different subsets of the data.

Question 11

When would you use precision over recall?

Answer 11

When false positives are more costly than false negatives.

Question 12

When would you use recall over precision?

Answer 12

When false negatives are more costly than false positives.

Question 13

What is ROC-AUC?

Answer 13

Area under the ROC curve; measures model’s ability to distinguish between classes.

Question 14

What is overfitting?

Answer 14

When a model learns noise in training data and performs poorly on unseen data.

Question 15

What is the sigmoid function?

Answer 15

s(z) = 1 / (1 + e^(-z)), e.g. s(0) = 1/(1+1) =0.5
- it maps a real valued number into the interval (0,1)
- useful for turning logits (raw prediction values) into probabilities in binary classification models