Introduction to Convolutional Neural Networks

Neural networks can be powerful for classification problems when you have unstructured data such as text or images. There is a type of neural networks that work particularly well on image data: Convolutional Neural Networks (CNNs). CNNs are a useful model for image recognition due to their ability to recognize visual patterns at varying scales. For image processing, densely connected networks can really grow very big if we have high resolution images. Therefore, CNNs are often preferred over densely connected networks. This blog is to give an introduction to the mechanism of CNNs.

Convolution

The essence of a CNN is the convolutional operation. A window is slided across the image based on a stride size. Padding can be used to prevent shrinkage and to make sure pixels at the edge of an image deserve the necessary attention. Each convolution then works to adjust the weights of the kernel through backpropagation during training. When using CNNs, you’re essentially changing your image through filters in every layer.

Pooling

Going back to the general architecture, max pooling is typically used between convolutional layers to reduce the dimensionality. After developing the convolutional and pooling layers to form a base, the end of the network architecture still connects back to a densely connected network to perform classification.