Convolutional neural networks

Question 1

Convolutional neural nets applications

Answer 1

• Computer vision
• Natural language processing
• Creating a robot that sounds like you
• Teach computers to play videogames
• Google maps use it to read street addresses and improve the maps

Question 2

model validation in keras

Answer 2

• trainining set
• validation set. Used to select the best parameters in the model. model.fit(validation_split) Modelcheckpoint
• test set. Set that has never been exposed to the model

Question 3

computer vision mlp vs cnn?

Answer 3

CNN

• CNN use layers that are sparsely connected, where the layer connections are informed by the 2d structure of the image matrix. This means fewer parameters than MLP
• CNN will accept the matrix as input and work with multidimensional data
• Each hidden node is responsible for gaining an understanding of a little part of the image or looking for patterns in just a part. Less prone to over fitting
• CNN are better. Good with images that are not easy or the number is small and hidden

MLP

• MLPs use a lot of parameters because only uses fully connected layers
• Only vectors as inputs. In an MLP you throw away all the 2d info when you flatten the image into a vector
• Every hidden node is responsible for gaining an understanding of the whole image or looking for patterns in the whole image

Question 4

convolutional neural nets

Answer 4

• Break the image into smaller pieces using a window of a defined height and width. Slide the window through the image.
• It is useful for each hidden node to share the same weights. This so that the different regions image share information
• Each window produces a node in the convolutional layer ( hidden layer) which connects to an input node
• Always add a relu function to the convolutional layer
• The weights for the inputs are the same for the window
• Convolutional layers detect regional patterns in an image.
• Always add a ReLU activation function to the Conv2D layers in your CNN. With the exception of the final layer in the network, Dense layers should also have a ReLU activation function.
• When constructing a network for classification, the final layer in the network should be a Dense layer with a softmax activation function. The number of nodes in the final layer should equal the total number of classes in the dataset.

Question 5

Stride and padding

Answer 5

• Stride is the size of the step of the filter when moving

- Padding is a surrounding none value info to allow steps

Question 6

Pooling layer

Answer 6

• Reduce the dimensionality of a convolutional layer
• Take a convolutional layer as Input
• Max pool is using a filter. Define the window and the stride
• Average is taking the average of all nodes in a feature map. Turning it into a vector

Question 7

How to take the learn understanding from a network and pass it to another?

Answer 7

• The technic is called transfer learning
• You can use a net trained for flowers, cats and dogs for recognizing cars or humans
• Conserve the layers that can recognize very simple shapes and eliminate layers that are very specific like the ones that can recognize dogs. Then add one or two more layers and train only them.
• Depending on both the size of the new data set and
  the similarity of the new data set to the original data set
  the approach for using transfer learning will be different.

Question 8

Types of layers in CNN

Answer 8

• input Layers
• Convolutional layers
• Pooling layers

Question 9

Pooling layers pancakes story

Answer 9

• Pancake is a convolution Layer that presents a feature map. Each pooling layer is a smaller pancake

Question 10

Convolutional layer

Answer 10

• It is a hidden layer
• Each node in the layer corresponds to a small part of the big image that is derived from a filter window
• Always apply a relu activation function to the convolutional layer
• Each Filter is designed to detect a particular pattern

Question 11

How to do learning transfer?

Answer 11

• Depending on both the size of the new data set, and
  the similarity of the new data set to the original data set
  the approach for using transfer learning will be different.
• There are four main cases:
  new data set is small, new data is similar to original training data
  new data set is small, new data is different from original training data
  new data set is large, new data is similar to original training data
  new data set is large, new data is different from original training data

Question 12

What is image augmentation?

Answer 12

• It is related to being able to recognize an object in an image even if it is translated to the left, right, up or down of the image.
• you can train a network to do this by giving it images with the objects translated or rotated
• keras imagedatagenerator