Parameters in Convolutional Network
It is a very important topic that is ignored by most beginners, they see the Parameters using the summary function but never try to alter them and analyze How it is affecting the accuracy of their Model. Because of this many problems arises. I will try to explain to you the consequences of a large number of parameters and ways to decrease the parameters in the Convolutional Network.
In this article, we will discuss
- Problems caused by a Large number of Parameters in Convolutional Network
- How to decrease Parameters of Convolutional Network to deal with those Problems
In this article, we’ll also learn the formula to calculate Parameters for each layer in Convolutional Neural Network.
Problems with a Large number of Parameters
Many Problems arises when we work with a Network having a Large number of Parameter. One of the problems comes with computational cost and execution time, both get increases in this situation. And this can’t be tolerated. In technical terms, it is called Time Complexity.
Let us take an example of the ANPR(Automatic Number Plate Recognition) system to understand this problem, We have installed it on a Traffic light to get the Vehicle Number of vehicles Violating Traffic laws if our model is taking few seconds to do so, Vehicle will pass on and our model fails. Here our model must take the least time to detect Vehicle numbers.
Another problem is the size of the model gets increases because of the increase in the number of parameters. Because of which space consumption increases. This is our job to make a model consuming the least space. In technical terms, it is called Space Complexity.
These two problems are enough to make us reduce the parameters of the Convolutional Network.
Decreasing Parameters of Convolutional Neural Network
- Replacing Dense and Flatten Layers with Global Average pooling
- Reducing the size of filters in Convolutional Layers https://towardsdatascience.com/understanding-and-calculating-the-number-of-parameters-in-convolution-neural-networks-cnns-fc88790d530d
I hope now you understood how important is it, to decrease the parameters of Convolution Neural Network to use it in production.
