Convolutional Neural Network

Question 1

How CNN learn visual features?

It uses __ __ in the input to inform the model.

Answer 1

It uses spatial structure in the input to inform the model.

Question 2

What is the idea of spatial difference?

It connect __ of input to __ in __ layers.

Answer 2

It connect patches of input to neurons in hidden layers.

Question 3

How does the patches in input layer connect with each other?

They connect each other by __ __.

Answer 3

They connect each other by sliding window.

Question 4

Filters input into patches. What is the operation called?

Con.. Op..

Answer 4

Convolution operation.

Question 5

What are the 3 steps of convolution operation?

1. Apply a filter - to extract __ __.
2. Use __ __ to extract __ features.
3. Spatially share __ across each filters.

Answer 5

1. Apply a filter - to extract local features.
2. Use multiple filters to extract different features.
3. Spatially share parameters across each filters.

Question 6

What does convolution produce?

It produce __ __.

Answer 6

It produce feature maps.

Question 7

What are the 3 steps to train CNN?

1. __: Apply filters to generate feature maps.
2. _____: ReLU
3. __: Downsampling operation

Answer 7

1. Convolution: Apply filters to generate feature maps.
2. Non-linearity: ReLU
3. Pooilng: Downsampling operation.

Question 8

In convolutional layers, what is the local connectivity?

The neuron in __ layer are computed as __ sum between patch __ weighted by __ patch. Each neuron only sees a __ patch.

Answer 8

The neuron in hidden layer are computed as weighted sum between patch input weighted by filter patch. Each neuron only sees a single patch.

Question 9

What is the computation of layer wise operation?

There are 6 steps.

1. Each inputs are neurons in patch of previous layers.
2. Apply __ __ __ (filters)
3. Do ___ operations
4. Add o__.
5. Add b__.
6. Activate with __.

Answer 9

1. Each inputs are neurons in patch of previous layers.
2. Apply matrix of weights (filters)
3. Do element-wise operations
4. Add outputs.
5. Add bias.
6. Activate with non-linearity

Question 10

What does pooling do?

Give 2

1. Reduce dimensionality of __ & increase dimensionality of __.
2. Preserve __ __.

Answer 10

1. Reduce dimensionality of features & increase dimensionality of filters.
2. Preserve spatial difference.

Question 11

Mini Conclusion to CNN: What are the 2 major parts of CNN?

1. F__ L__
2. C__ P__

Answer 11

1. Feature Learning
2. Class Probabilities

Question 12

Object Detection involves with detecting objects into boxes to classify each boxes into their own class.
How to effectively detect object?

The naive approach is bad cause it detect too many boxes.

…-…. and Faster …-…

Answer 12

R-CNNs and Faster R-CNNs

Question 13

What are the 2 operations for semantic segmentation?

Answer 13

Downsampling and upsampling

Downsampling broke down the image into pixels of classes.
Upsampling build these pixels back into an image.