Quiz 2

Question 1

Receptive fields

Answer 1

Each node only receives input from a 𝐾1×𝐾2 window (image patch).

The region from which a node receives its input from is called a receptive field.

Question 2

Shared Weights

Answer 2

Nodes in different locations can share features.

Uses the same weights/parameters in the computation graph.

• Reduce parameters to (𝐾1×𝐾2+1)
• Explicitly maintain spatial information

Question 3

Learning Many Features

Answer 3

Weights are not shared across different feature extractors.

Reduce parameters to (𝐾1×𝐾2+1)∗𝑀 where M is the number of features to be learned

Question 4

Convolution

Answer 4

In mathematics a convolution is an operation on two functions f and g producing a third function that is typically viewed as a modified version of one of the original functions, giving the area of overlap between the two functions as a function of the amount that one of the original functions is translated.

Question 5

T or F: Convolutions are linear operations

Answer 5

True

Question 6

What are CNN hyperparameters

Answer 6

1. in_channels(int): Number of channels in the input image
2. out_channels(int): Number of channels produced by the convolution
3. Kernel_size(int;tuple): Size of convolving kernel
4. stride (int;tuple;optional): denotes the size of the stride used by the convolution (default is 1)
5. padding (int;tuple;optional): Zero padding added to both sides of the input (default is 0)
6. padding_mode (string):
   ‘zeros’,’reflect’,’replicate’,’circular’ (default ‘zeros’)

Question 7

Output size formula for a vanilla convolution operation

Answer 7

Specifically:

(N - F + 2P) / S + 1 = output_dim_1 * output_dim_2 * N_channels

N: Input dimension
F: Filter dimension
P: Padding
S: Stride
1: Bias term

Question 8

“valid” convolution

Answer 8

where the kernel fits inside the image

Question 9

T or F: Larger the filter the smaller the shrinkage

Answer 9

False. Larger filter = larger shrinkage.

Question 10

“Same” convolution

Answer 10

zero-padding the image borders to produce an output the same size as the raw input

Question 11

CNN: Max pooling

Answer 11

For each window, calculate its max.

Pros: No parameters to learn.

Question 12

CNN: Stride

Answer 12

Movement of the convolution layer.

Question 13

CNN: Pooling layer

Answer 13

Make the representations smaller and more manageable through downsampling.

Only pools width and height, not depth

Question 14

CNN: Cross-correlation

Answer 14

Takes the dot product of a small filter (also called a kernel or weights) and an overlapping region of the input image or feature map.

Doesn’t flip the convolution layer.

Question 15

CNN: T or F - Using a stride greater than 1 results in loss of information.

Answer 15

True. Stride > 1 implies jumping over some pixels.

Question 16

CNN: Output size of vanilla/valid convolution vs full convolution

Answer 16

Vanilla: m-k+1
Full: m+k-1

Question 17

CNN: Benefit of pooling

Answer 17

Makes the representation invariant to small changes in the input

Question 18

CNN: Full convolution

Answer 18

Enough zeros are added to the borders such that, every pixel is visited k times in each direction. This results in an image of size m+k-1.

Full = Bigger size than original

Question 19

Sigmoid

Answer 19

Min=0; max=1

Output is always positive

Saturates at both ends

Gradients vanish at each end (converging to 0 or 1 - gradient approaches zero)

Always positive

Computationally complexity high due to exponential term

Question 20

tanh

Answer 20

min=-1; max=1; and we note that is centred
Saturates at both ends (-1,1)
Gradients: vanish at both ends ; always positive
medium compexity as tanh is not as simple as say multiplication

Question 21

ReLU

Answer 21

Min=0, Max= ∞; always positive
Not saturated on the positive side
gradients: 0 when X <= 0 (aka dead ReLU); constant otherwise (doesn’t vanish which is good)
Cheap: doesn’t come much easier than max function

Question 22

T or F: ReLU is differentiable

Answer 22

Technically no, but only at zero.

Question 23

Initialization: What happens if you initialize close to a bad local minima

Answer 23

Poor gradient flow

Question 24

Initialization: What happens if you initialize with large activations

Answer 24

Reach saturation quickly

Question 25

Initialization: What happens if you initialize with small activations

Answer 25

Be in the linear regime or close to it in the nonlinear space, and you will have a strong gradient to learn from

Question 26

Initialization: What happens if you initialize all weights with a constant

Answer 26

All learns the same thing.

Question 27

Initialization: Common practice

Answer 27

1) Random sample from small normal distribution 𝑁(𝜇=0,𝜎=0.01) 2) Random sample from uniform distribution

Question 28

Initialization: Why are equal (in terms of sampling from distribution), small weights preferred

Answer 28

No a priori reason why some weights should be greater

Question 29

Initialization: T or F - Deeper networks are less sensitive to initialization

Answer 29

False. More sensitive with deeper network because activations increasingly get smaller.

Question 30

Initialization: Fan-in Fan-out rule

Answer 30

Maintain the variance at the output to be similar to that of the input. Keeps each layer's variance the same.

Question 31

Optimization: Issues that hinder optimization

Answer 31

Noisy gradient estimates (due to taking MiniBatches) Saddle points Ill conditioned loss surface, where the curvature is high in one direction but not the other

Question 32

Optimization: Loss surfaces that can cause problems

Answer 32

Local minima plateaus saddle points, a point that is a min in one axis but a max in another

Question 33

Optimization: Momentum

Answer 33

Overcomes plateaus by adding exponential moving average of the gradient. Helps move off of areas with low gradients.

Question 34

Optimization: Nesterov Momentum

Answer 34

Calculates gradient AFTER applying momentum term

Question 35

Optimization: How to use Hessian

Answer 35

Use 2nd order derivatives to get information about the loss surface

Question 36

Optimization: Condition number

Answer 36

The ratio between the smallest and largest eigenvalue of a hessian Tell us how different the curvature is along different dimensions

Question 37

Optimization: General idea of techniques like Adam

Answer 37

Apply per-parameter learning rates

Question 38

Optimization: Adagrad

Answer 38

Adapts the learning rate for each parameter based on the historical gradients Larger gradients have a rapid decrease in LR, while those with small gradients get a slower decrease. Pro: Works well in a gently sloped parameter space. Con: Can prematurely make learning rate too small.

Question 39

Optimization: RMSProp

Answer 39

Takes AdaGrad but replaces calculation by an exponential moving average

Question 40

Optimization: Adam

Answer 40

Like Adagrad / RMSProp but includes momentum terms

Question 41

Regularization: L1 norm

Answer 41

Applies the a sign function to the weights in addition to regularization term. Results in sparse parameters 𝐽=𝛼 sign( 𝑊^{𝐽−1}+𝐽(𝑊^{𝐽−1}) )

Question 42

Regularization: L2 norm

Answer 42

Applies a regularization term to the weights at each update 𝐽=𝛼 𝑊^{𝐽−1}+𝐽(𝑊^{𝐽−1})

Question 43

Regularization: Dropout

Answer 43

Dropout is a technique in which a set of parameters are randomly masked (ie a matrix of 1/0's) to prevent a subset of params from learning. Makes the model less reliant on specifically effective parameters

Question 44

Regularization: What needs to be done with dropout during inference

Answer 44

1) Scale outputs or weights by the masking probability "p" W_test = W * p 2) Scale by 1/p during training.

Question 45

Batch norm: Diff between batch vs layer norm

Answer 45

Batch: Normalizes activations along the batch dimension Layer: Normalizes activations along the feature (channel) dimension for each data point in the mini-batch

Question 46

Batch Norm: Definition

Answer 46

Normalizes the activations of a layer across a mini-batch of data, which helps stabilize and speed up training by reducing internal covariate shift.

Question 47

Batch Norm: How to use batch norm during inference

Answer 47

Take running average taken during training.

Question 48

Batch norm: Pros

Answer 48

Improves gradient flow Allows higher learning rates Reduces dependence on initialization Differentiable

Question 49

Batch norm: Cons

Answer 49

Sufficient batch sizes must be used to get stable per-batch mean/variance.

Question 50

Batch norm: Where to apply in network

Answer 50

Right before activation

Question 51

Batch norm: T or F - Batch norm is useful for linear networks

Answer 51

False. This is because we have normalized out the first- and second-order statistics, which is all that a linear network can influence

Question 52

Batch norm: T or F - Batch norm is useful for deep networks

Answer 52

True. In a deep neural network with nonlinear activation functions, the lower layers can perform nonlinear transformations of the data, so they remain useful.

Question 53

Batch norm: T or F - The bias term should be omitted during batch norm

Answer 53

True. The bias term should be omitted because it becomes redundant with the β parameter applied by the batch normalization reparametrization.

Question 54

Batch norm: T or F - For CNNs, you should use different mean/variance parameter values for each layer

Answer 54

False. It is important to apply the same normalizing μ and σ at every spatial location within a feature map, so that the statistics of the feature map remain the same regardless of spatial location.

Question 55

Initialization: Variance of ReLU

Answer 55

𝑁(0,1)×sqrt(𝑛_j / 2)

Question 56

Batch: Batch gradient descent

Answer 56

Network goes through the forward/backward pass using the entire dataset. Pro: Provides deterministic updates. Con: Computationally expensive.

Question 57

Batch: Stochastic gradient descent

Answer 57

Network goes through f/b using only one example. More noisy updates, but can have regularization effect by pushing model out of local minima. Doesn't take advantage of parallelization

Question 58

Batch: Mini-batch gradient descent

Answer 58

Takes N samples from the empirical distribution and runs f/b pass. Benefits from parallel processing. More stable than SGD.

Question 59

Supervised pretraining

Answer 59

Model is initially trained on a related task that has labeled data (supervised learning) before fine-tuning it on the target task of interest. Knowledge gained during the initial training phase can serve as a useful starting point for the model when tackling the target task.

Question 60

Formula to calculate number of parameters and bias

Answer 60

K (F1 * F2 * D + 1) K: Number of kernels F: Filter dimensions D: Number of input channels