L9 - Computer Vision and CNN

Question 1

1. What is the aim of Computer Vision?

Answer 1

1. Use ANN’s to mimc what the brain is doing when humans use their eyes.

Question 2

2. What is a naive approach to using a deep neural network for image classification?

Answer 2

1. Stack image pixels into a vector of binary format
   
   2. Learn the vector

Question 3

3. What are the issues of stacking pixels and running them through a DNN for image classification?

Answer 3

1. High-dimensional and doesn’t scale
   
   2. Not robust to small changes

Question 4

4. What did Hubel and Wiesel notice? What neural network did this lead to?

Answer 4

1. Different neurons fire when vision target changes
   
   2. Convolutional Neural Network

Question 5

5. What 3 practicalities are CNN’s based on?

Answer 5

1. Reduces number of input nodes
   
   2. Tolerates small pixel changes with no impact on classification ability
   3. Takes advantage of pixel correlations on complex images

Question 6

6. What is a filter in CNN? How is it learned?

Answer 6

1. Filter is a low dimension kernel learned via back propagation on the original image.

Question 7

7. How is the filter applied to the image?

Answer 7

1. Place the filter ‘over’ the original image, and perform Dot Product between the filter and the original image.
   
   2. Add a bias term
   3. Returns a value of either 0 or 1
   4. Value is then added to a feature map
   5. Move filter one position and repeat
   6. Stop when feature map is complete

Question 8

8. What does convolution mean?

Answer 8

1. The process of moving the filter over the image and adding the value to the feature map. Hence the name CNN.

Question 9

9. What is the difference between pre-1980 CNN’s and post-1980 CNN’s?

Answer 9

1. Post 1980 introduced Deep Convolutional Neural Networks with nested feature maps.

Question 10

10. Describe the Deep Convolutional Neural Network process…

Answer 10

1. original 64 x 64 image, with a 5x5 filter.
   
   2. Perform convolutions to obtain 60x60 feature map
   3. 60x60 feature map is used as input for the next iteration.
   4. Process repeats, creating a deep nesting of feature maps.
   5. Each iteration creates a more accurate feature map

Question 11

11. In a Deep CNN, what are FC Layers?

Answer 11

1. Fully connected -> Everything connects to everything in each layer

Question 12

12. In a Deep CNN, what is a Soft Max?

Answer 12

1. Deep CNN connects down to 10 neurons

Question 13

13. What are 3 limitations to Deep Learning?

Answer 13

1. Computationally expensive -> Needs GPU
   
   2. Hard to interpret -> Making them Black Box like
   3. Need lots of training data