[Terms] CNN Basics

Convolutional Neural Network is mostly used in image data.

 

Because the data is in a grid a discreate convolution is used. The filter is moved over the data and the overlapping cells are multiplied. Then the multiplied values are added to get the value for the data.

 

This filter can be used for data other than images. Because the filter has its dimensions the CNN can be used in sequential data to get the relationship between the sets. This is still discouraged because in CNN only the data till the filter dimension can be considered while in RNN models all the previous data is considered.

 

There are lots of different filters that can be used to get the data we want. For images LPF(Blur), HPF(Edge Detection) are the most common ones.

https://en.wikipedia.org/wiki/Kernel_(image_processing)

 

Here are some of the parameters passed…

• Max/average pooling: The filter is moved across the data, but this time we find the max/average value in each overlapped data. If a pooling of 2x2 is used, then the overall image’s dimension is reduced to half its sides.
• Padding: Before the data is used in a convolution/pooling the data is surrounded by extra values. This is used to manipulate the output data’s size. If a padding of n is used, then the overall image’s dimension is increased by 2n.
• Stride: This is the distance in which the filter is moved per operation. If a stride of n is used, then the overall image’s dimension is reduced by 1/n to its sides.