CNN

Question 1

IMAGENET dataset

Answer 1

Dataset about daily use materials and animals
2272273
* 1,281,167 training images,
* 50,000 validation images, and
* 100,000 test images
* 1000 classes

Question 2

CNN

Answer 2

Special kind of neural network for processing data that has a grid-like topology, like time series data (1D) or image data (2D).

CNNs consists of: -
1. Convolution layer
2. Pooling layer
3. Fully connected layer (ANN)

Question 3

Why not use ANN on image data?

Answer 3

1. Computational complexity (Large no. of pixels)
2. Overfitting
3. Loss of spatial arrangement (Since, 2D image is converted to 1D layer)

Question 4

How does CNN work on image data?

Answer 4

Initial convolutional layer extracts primitive features (edges).
Going further in network, more complex features are extracted.

Question 5

Greyscale image

Answer 5

B/W
Single channel
Values between [0-255]

Question 6

Colored Image

Answer 6

RGB
Three channel

Question 7

What is convolution?

Answer 7

Convolution is element-wise matrix multiplication, where kernel (filter) is multiplied with the input pixels to get the feature map.

The process of detecting features in an image is called convolution.

Question 8

How is the value of filter (kernel) is decided?

Answer 8

Initialized with random value.
Decided during backpropagation

Question 9

What is filter?

Answer 9

A matrix of weights that slides over the input pixels, perform element wise multiplication to give a single output pixel.

Question 10

What is padding?

Answer 10

Contribution of edge pixel is less to form the output than the central pixels. In order to make them equal, we use padding.

Question 11

What is stride?

Answer 11

Stride decide how our weight matrix should move in the input, i.e. jumping one step or two.

Question 12

Valid padding

Answer 12

No padding

Question 13

Same padding

Answer 13

Automatic padding so that size of input image is same as feature map

Question 14

Formula to find the output after convolution

Answer 14

[n + 2p - f ]/s + 1

Question 15

Why are strides required?

Answer 15

1. Extract only high level features
2. Limit feature; helps reduce complexity

Question 16

Why is pooling required?

Answer 16

This is because convolution has:

1. Memory issue
2. Translation variance

Though, increasing the stride will address the memory issue but translation variance problem will not be solved by stride.

Pooling down sample the feature map.

Question 17

Translation Invariance

Answer 17

The ability to ignore positional shifts, or translations, of the target in the image.

Question 18

Type of pooling

Answer 18

1. Max pooling
2. Avg pooling
3. Global pooling (Global max & Global avg)

Question 19

Advantage of pooling

Answer 19

1. Reduced image size (due to down sampling)
2. Translation invariance
3. Enhanced feature (Only in max pooling)

Question 20

Disadvantage of pooling

Answer 20

1. Not suggested for Image segmentation tasks
2. Loss of information

Question 21

ANN vs CNN

Answer 21

Similarity: -
1. Input*weights + bias; both works in the same way. In case of CNN, weights means filters’ weights.

Differences: -
1. No. of learnable parameter do not depend on input in CNN

Question 22

How to reduce overfitting in CNN model?

Answer 22

1. Add more data
2. Data Augmentation
3. L1/L2 Regularization
4. Batch Normalization
5. Dropout

Question 23

Why do we need data augmentation?

Answer 23

1. To generate more data
2. To reduce overfitting (Increase generalization of image)

It includes Image rotation, scaling, flip, zoom,

Question 24

Why do we need pretrained models?

Answer 24

1. Absence of labeled data
2. CNN is computationally expensive to train

Question 25

What is pooling

Answer 25

Also called down sampling, as it carries out dimensionality reduction. The feature parameters in the input are reduced to only the necessary parameters to reduce complexity and improve the network’s performance. It also helps avoid the problem of overfitting.

Pooling operation involves sliding a two-dimensional filter over each channel of feature map and summarizing the features lying within the region covered by the filter.

Question 26

Feature map

Answer 26

Output of the convolutional layer, a numerical representation of the image, which is used to identify patterns from the image.

Question 27

ResNet

Answer 27

ResNet stands for residual network, which refers to the residual blocks that make up the architecture of the network.

Question 28

Residual connections

Answer 28

Skip connections, also known as residual connections are implemented by adding the output of an earlier layer to the output of a later layer.

They allow for the preservation of information from earlier layers, which helps the network to learn better representations of the input data and solve the problem of vanishing gradients.

Question 29

Key Features of ResNet-50

Answer 29

• ILSVRC’15 classification winner (3.57% top 5 error ResNet-152)
• Has other variants also (with 35, 101, 152 layers)
• Every ‘residual block‘ has two 3×3 convolution layers
• No FC layer, except one last 1000 FC softmax layer for classification
• Global average pooling layer after the last convolution
• Batch Normalization after every convolution layer
• SGD + momentum (0.9)
• No dropout used