Module 5

Question 1

Binary classification

Answer 1

Only two possible classes

Question 2

Multi-class classification

Answer 2

Multiple classes

Question 3

Multi-label classification

Answer 3

Each input can belong to more than one class

Question 4

Batching

Answer 4

• combining vectors of several data points into one matrix
• improves speed
• reduces noise

Question 5

Backpropagation

Answer 5

Propagating the gradients backwards through the network layers

Question 6

Low learning rate

Answer 6

Model will take forever to converge

Question 7

Learning rate too high

Answer 7

Keep stepping over the optimal values

Question 8

Learning rate decay

Answer 8

Reducing the learning rate by a factor

Question 9

Weight initialisation

Answer 9

• zeros

- draw randomly from N(0,1)

Question 10

Network capacity

Answer 10

Correlation between the capacity of a neural network and its ability to overfit

Question 11

Network underfitting

Answer 11

Increase the number of neurons/layers

Question 12

Network overfitting

Answer 12

Lowering the number of neurons/layers

Question 13

Regularisation

Answer 13

Adding some information/constraints to stop the model from overfitting

Question 14

L1 regularisation

Answer 14

Adding the absolute weights to the loss function

Question 15

L2 regularisation

Answer 15

Adding the squared weights to the loss function

Question 16

Dropout

Answer 16

training: randomly set some neural activations to zero
testing: use all neurons but scale the activations