Understanding Filters in Convolutional Neural Networks: A Comprehensive Guide

Understanding Filters in Convolutional Neural Networks: A Comprehensive Guide🔎🧠📊

In this article, we delve into the fascinating concept of filters within Convolutional Neural Networks (CNNs). We will explain how filters are used to extract features from input data and create feature maps.

Why Filters?

In a CNN, each cell in a convolutional layer receives information from a group of pixels or cells in the previous layer. But how does this happen? How is a cell able to gather information from numerous pixels? We have 25 pixels here and our cell can only hold one value representing all these pixels. So we need a way to convert these 25 values into one value.

The Role of Filters

Filters are matrices with the same dimensions as the receptive field (e.g., if the window is 5x5, the filter will also be 5x5). A filter multiplies each pixel value by the corresponding filter value and adds up all these products to generate a single number, representing the information in those pixels.

Determining Filter Values

The values in the filters are not manually chosen. During training, our network learns these values itself.

Demystifying Filters with Examples

To illustrate how filters function and highlight specific features, we'll use a 5x5 input image with pixel values of 0.1 and a 3x3 filter. Observe this filter: it resembles a cross, with diagonal values being 1 and the others 0.

With a stride of 1, using this filter, we get the output as shown in the GIF below.

By multiplying the pixel values with their respective filter values and summing up these products, we obtain a number that represents the information within these 25 pixels. Now, to demonstrate how filters work, I have taken a 5x5 input image with 0.1 type pixel values and a 3x3 filter.

Feature Maps

The output after applying a filter is called a feature map. Each feature map highlights a specific feature of the input image. To better understand this, let's examine some examples:

• 
  Using a vertical filter (middle column is 111 and side columns are 0.00), we transform the image to highlight vertical white lines.
  
• 
  Using a horizontal filter (middle row will be 111 and top/bottom rows contain 0s), we obtain an image where horizontal white lines are emphasized.
  

Multiple Filters for Multiple Features

To create various feature maps with different features, we apply multiple types of filters. Each convolutional layer becomes a bundle of feature maps, each highlighting specific features. This is crucial because in the next layer, each cell will process information from all feature maps in the previous layer to identify more complex features.

Stay Tuned!

In the next article, we'll discuss channels in CNNs and explore how they impact the output layer.

FAQs

1. 
   What is a filter in CNN?
   Answer: A filter is a matrix used to extract features from the input data and create feature maps in a Convolutional Neural Network.
   
2. 
   How does a filter work?
   Answer: Filters multiply each pixel value by the corresponding filter value and sum up these products to generate one number, representing the information within those pixels.
   
3. 
   What happens if we use different filters in a CNN?
   Answer: Using various filters allows us to create multiple feature maps, each highlighting a distinct feature of the input data.
   
4. 
   Can filters learn their values by themselves during training?
   Answer: Yes! The network learns filter values as part of the training process.
   
5. 
   What are feature maps in CNNs?
   Answer: Feature maps are the output after applying a filter to the previous layer's data, highlighting specific features within the input image.