What is a Convolutional Neural Network?

What is a Convolutional Neural Network?

A Convolutional Neural Network (CNN) is a Deep Learning algorithm used in image processing by computers. CNNs are able to classify images by distinguishing important image features from unimportant image features. The structure of CNNs is based on a biological model: the visual cortex (visual cortex) of the human brain. 

How does a CNN work?

The following is a simplified explanation of a CNN without mathematical details: A CNN has three types of layers: Convolutional Layer, Pooling Layer and Fully Connected Layer.

Convolutional layers are responsible for recognizing and processing the features of an image, such as shapes or edges. This is done by analyzing the image from different filters. The filters have a certain pixel size and gradually scan the image for its features. This can be thought of as using a magnifying glass to scan the image from left to right as well as from top to bottom. The results of the scanning process are recorded in a result matrix. The result matrix is then scanned by a smaller filter and these results are also recorded in a result matrix. This process is repeated several times so that the image is analyzed down to the smallest detail. 

The very precise analysis of the image by the Convolutional Layer produces a large amount of data, much of which is unimportant for the actual image processing. The decision as to which data is important and which can be discarded is made in the Pooling Layer. The pooling layer thus ensures a dimensional reduction and a condensation of the information on the scanned features. This is crucial because it significantly reduces the amount of data and thus allows the data to be processed much faster.

After some repetitive units consisting of Convolutional Layer and Pooling Layer, the Fully Connected Layer links the results together and is thus able to classify images.

In case you want to see how each layer looks and works: Ryerson University has developed an interactive 2D visualization of a CNN that can recognize handwritten numbers: Ryerson University

At claimflow, we use CNNs to classify incoming documents.

Sources (translated): Towards Data Science and Towards Data Science