Machine Learning NanoDegree

Question 1

apply bias and variance to ( underfitting || overfitting )

Answer 1

bias - underfitting, variance - overfitting

Question 2

define the harmonic mean for (x, y)

Answer 2

2xy/x+y

Question 3

what is the f1 score

Answer 3

the harmonic mean of precision and recall - raised a flag if any of the values are small

Question 4

what is precision

Answer 4

the percent of labeled positives that are actually positive

Question 5

what is recall

Answer 5

the percent of total positives that are label positive

Question 6

What is fbeta score?

Answer 6

an f1 score that allows biasing towards either precision or recall. beta = 1 = harmonic mean, beta > 1 tends towards recall, 0 < beta < 1 tends towards precision

Question 7

what is an ROC curve and how do you interpret it?

Answer 7

close to 1 is good, 0.5 is random,

Question 8

What is r2 score and how do you interpret it?

Answer 8

the difference between a regression model and the simple averaging of all the points. close to 1 is good, close to 0 is bad

Question 9

What is the point of having a bias node in a NN layer?

Answer 9

To provide the constant or intercept.

Question 10

What is the ‘perceptron trick’ to get a line to move closer to a point?

Answer 10

subtract the the point vector (plus one for bias) time the learning rate from the linear equation if the point is negative labeled positive, add if the point is positive labeled negative.

Question 11

What is the formula for multi-class entropy?

Answer 11

• sum(for i in p){ p[i] * log2(p[i]) }

Question 12

what does ‘naive’ refer to in naive bayes?

Answer 12

assuming that all variables are independent.

Question 13

a function must be ___ not ___ in order to be optimized

Answer 13

continuous, discreet

Question 14

describe l2 regularization, including its alternate name

Answer 14

also called ridge regression, l2 regularization adds the square of the coefficients to the cost function, perhaps scaled by lambda. This works to penalize the model for being too complex and reduce overfitting.

Question 15

describe l1 regularization, including its alternate name

Answer 15

also called lasso regression, l1 regularization adds the absolute value of the coefficients to the cost function, perhaps scaled by lambda. This works to penalize the model for being too complex and reduce overfitting. Reduces less important features to 0 and thus may be suitable for feature engineering.

Question 16

What is a polynomial kernal and what does degree refer to?

Answer 16

A polynomial kernal projects a two dimensional function into 5 dimensions by adding terms x^2 , xy, and y^2. Higher degree polynomials add more exponents and combinations and therefore higher dimensions.

Question 17

is softmax for n=2 the same as sigmoid activation ?

Answer 17

yes

Question 18

how is softmax defined?

Answer 18

for i, softmax = ei / sum(e0…e**n)

Question 19

What is cross entropy ?

Answer 19

-log(P)

Question 20

What is the chain rule?

Answer 20

the partial derivative of a series of composed functions is equal to the product of the partial derivative of each of the functions

Question 21

What is a monotonic function?

Answer 21

A function which is either entirely non-decreasing or non-increasing.

Question 22

What is early stopping?

Answer 22

Stop training when the cross - val error starts to increase

Question 23

how do you use a nn for regression instead of classification?

Answer 23

remove the final activation function and let it return the result of the last layer.