Learning Algorithms

Question 1

• Regression vs Classification

Answer 1

• Continuous vs discrete

Question 2

• Parameter vs Hyperparameter

Answer 2

• Parameter – properties learned by model during training (DT threshold value or NN connection weight)
• Hyperparameter – this you choose before learning process (Dt: branch amounts or NN hidden layers and size)

Question 3

• Strategy for automatic selection of feature values on branch splitting

Answer 3

Choose a statistical term to use as the threshold (mean, median)

Question 4

• What is black box in learning algorithm

Answer 4

• Multiple hidden layers or nodes that are not easily interpretable

Question 5

What is meant by generalization and how well do DT and NN perform at this?

Answer 5

• DT: based on tree depth and pruning, prone to overfit not good at generalising
• NN: based on complexity and regularization techniques, more good to generalising, less likely to overfitting.

Question 6

We used different names for training and loss function in physics based modelling, what are they?

Answer 6

• Calibration and objective function

Question 7

• How to extend classification models to predict categories?

Answer 7

Create a binary classifier to compare the features against each other