Tut 5

Question 1

LReLU

Answer 1

f(x) = x if x >= 0; ax if x < 0
Where a is some predetermined constant

Question 2

PReLU

Answer 2

f(x) = x if x >= 0; ax if x < 0
Where a is some learned parameter

Question 3

Def Feature map

Answer 3

The output produced by a single mask (before or after the application of the activation function, or pooling) in a convolutional layer of a CNN.

Question 4

Def sub sampling

Answer 4

Process of reducing the width and height of a feature map by pooling (also the process of reducing the resolution of an image)

Question 5

Def adversarial example

Answer 5

A sample that has been manipulated so that it is classified differently than the original sample

Question 6

Def regularisation

Answer 6

Any method intended to avoid overfitting and improve generalisation

Question 7

Def dropout

Answer 7

A method to reduce over fitting that works by giving an activation of 0 to a random sample of neurons in a layer at each iteration during training

To avoid specific neurons being associated to specific data samples

Question 8

Alernative way to calculate Var[x]

Answer 8

Var[X] = E[ (X - E[X])²]

Question 9

Formula for calculating output height/width of convolutional layer

Answer 9

Output_height = 1 + (input_height - mask_height + 2 x padding)/stride

Question 10

What does 1 x 1 convolution do

Answer 10

Creates weighted sum of input channels at each location
Flattens channels into 1