VL VIII - Convolutional Network

What does FC mean

Fully connected

Why not use FC forever

• No structure

• Network cannot be arbitrarily complex

• Optimization becomes hard

• Performance plateaus

Why use CNN

• Layer with structure

• Weight sharing (same wights for different parts of image)

What does CNN mean

Convolutional Neural Network

What to do at Boundaries of Filters

1. Shrink

2. Pad (adding a number at the edge)

Whats the impact of the depth dimension

The filter has to match the dimension of the input

How to convolve filter with an image

• apply filter at each location

• dot products

How to the filter weights differ with a larger image (RGB)

they dont, they stay the same independant of image size

what changes with Input size (CNN)

Output gets larger as well

Whats the output of a conv layer

feature map/activation map

how do the filter in the same conv layer differ

They have different weights values, but the same dimension and amount of weights

How is a convolution layer defined

• Filter width and height (depth given thorugh input)

• Number of different filters

• each filter captures a different image characteristic (e.g. circles, squares, etc.)

Whats stride in regards to CNN

• The step witdh at which the filter is applied and gives an output

Whats an illegal stride

When the filter does not fully fit

how to calculate output size CNN no padding

why use padding CNN

• sizes get to small too quickly

• corner pixel is only used once

Which padding is the most common

Zero padding

Types of convolution (Padding)

• Valid convolution (no padding)

• Same Convolution (set padding (F-1)/2)

why use same convolution

• size of the output stays the same as the input

how to calculate output size CNN with padding

How to calculate Parameter for each filter

Filter dimemsion x depth of input

(5 × 5) x 3

How to calculate number of Parameters per layer

Number of filter parameters x number of filters

You are given a convolutional layer with 4 filters, kernel size 5, stride 1, and no padding that operates on an RGB image

What are the dimension and the shape of its weight tensor

[3,4,5,5]

different types of pooling

• max pooling

• average pooling

• min pooling

difference Conv Layer and Pooling Layer

• Feature Extraction: computes feature in given region

• Feature selection: picks the strongest activation in region

R

How does the pooling layer influence the conv layer

Pooling layer wants to extract promising feature, so conv layer hast to be optimized so pooling can extract set feauture

hyperparameters of pooling

• Stride

• Spatial filter extent

what does FC do at the end of CNN

• make final decision with extracted feature of Conv layers

• one or two FC layers

Difference Conv to FC

• restrict degrees of freedom

  • FC is brute force

  • Conv are structured

• Sliding window with the same filter parameters to extract image features

  • concept of weight sharing

  • Extract same feature independant of location

Whats the receptive field

Spatial extent of the connectivity of a conv filter

When does the receptive field grow

with layer and filter kernel size

How is complexity achieved in CNN

More layers and easy filters