4404

Question 1

What is machine learning

Answer 1

Machine learns from exp E with respect to class of task T and performance measure of P. Performance P at tasks in T improves with exp E

Question 2

Machine learning branches

Answer 2

Supervised | Unsupervised
–
Regression, Classification |
Regression for continous data and predictions, Classification for discrete data and predictions

Question 3

Vectorization

Answer 3

use vectors and matrices to represent data and feed into an algorithm. This is called vectorization

Question 4

Cause of underfitting
Solution to underfitting

Answer 4

Cause:
Occurs when model is too simple to learn underlying structure of data
Solution:
Select a more complex model
Feed better features to learning alg
reduce regularization

Leads to high bias

Question 5

Cause of and solutions for overfitting

Answer 5

Cause:
Too complex of a model compared to size and noise of data
Solutions:
Use model with fewer parameters (linear vs high degree polynomial)
Constraint the model through regularization
Use more training data
Fix data errors and remove outliers

Question 6

Cost/Performance Y-axis with respect to Jtrain and Jval meanings

Answer 6

Lower cost represents less errors in predictions
Higher cost reps more errors
Relation between Jtrain and Jval represent how well a model is generalizing to the data

Question 7

Generalization

Answer 7

Ability of a model to make predictions on unseen data

Question 8

How to mitigate overfitting

Answer 8

increase regularization parameter
increase data samples
use fewer features
early stopping/fewer epochs for training

Question 9

How to mitigate underfitting

Answer 9

decrease regularization parameter
use a more complicated model/ more features

Question 10

Regularization

Answer 10

Prevent overfitting by adding a penalty to model complexity
Discourage model from fitting the training data to closely | includes noise and outliers

Question 11

Regularization parameter

Answer 11

Term added to cost function to train the model. Controls trade-off between fitting and keeping parameters small.
Larger param penalizes larger coefficients more strongly
Lower param allows model to fit training data more closely

Question 12

L1 Regularization (Lasso)

Answer 12

Adds a penalty to the absolute value of the coefficients
encourages sparsity in the model, driving some coefficients to zero selecting a simpler model with fewer features

Question 13

L2 Regularization (Ridge)

Answer 13

Adds a penalty equal to the square of the coefficients
encourages small coefficients, but not necessarily 0.

Question 15

Linear Classification

Answer 15

Perceptron is a linear supervised binary classification alg that can apply when classes are linearly separable.
If training data can be separated by linear decision rule, they are linearly separable.

Question 16

Hyperplane

Answer 16

A decision boundary separating different classes.

Question 17

Support vectors

Answer 17

Data points closest to the hyperplane. Used to maximize distance between hyperplane and nearest data points from either class.

Question 18

Linear SVM

Answer 18

When datapoints are linearly separable, SVM finds a linear hyperplane to separate the classes

Question 19

Non-linear SVM

Answer 19

When data is not linearly separable, SVM uses tech ‘kernel trick’ to transform data into higher dimensional space where a linear hyperplane can be used

Question 20

Logistics regression

Answer 20

Uses sigmoid function. Used to classify data into two distinct classes

Question 21

Sigmoid function

Answer 21

Maps any real-valued number into the range of 0, 1 | y E [0,1]
Output interpreted as the probability of that input belonging to the positive class