Virtusa Launches Helio to Help Enterprises Harness GenAI to Accelerate Outcomes | Read more

Digital Themes

Convolutional Neural Network

What is convolutional neural network?

In deep learning (neural nets), convolutional neural networks (CNN)s, also known as ConvNets, are specialized artificial neural networks that can “see” the world as humans perceive it. These algorithms take an input image, then assign importance to various objects within that image to differentiate one object from the other. Convolutional neural networks are part of the field of computer vision, which is the study of how computers use artificial intelligence (AI) to achieve a higher understanding of digital imagery or video.

How does a computer implement natural language processing to perceive the world? CNN architecture mimics the organization of the visual cortex, the area of the brain that processes visual information, as well as the connectivity of neural pathways in the human brain. Convolutional networks consist of fully-connected layers of artificial neurons or mathematic functions that can take pixel values and identify visual features of an image by creating activation maps (maps that highlight the features of an image). An activation map of the final convolution layer in the CNN generates a set of values used to assign a class to an input image.

More specifically, the base layer, or convolutional layer of a convolutional neural network establishes learning parameters, followed by pooling layers, which apply a given function or convolution operation to reduce spatial size of an image. Outputs from a pooling operation build on each other by passing through an activation function to create feature maps, extracting increasingly complex features along the way. Max pooling calculates maximum values of each feature map that results in movement from the dissemination of corners and edges, to that of more complex subjects such as the human face.

There are many aspects of conventional neural networks that are advantageous to their application.

  • Artificial intelligence (AI). Unlike a regular neural network, convolutional neural networks are able to identify and capture relevant features from an image or video by mimicking the human brain. As a result, CNNs consistently outperform neural networks on image recognition-related tasks.

  • Precise filter size. Convolutional neural networks take the guesswork out of determining filter size for general task feature extraction. The training of a CNN network involves the training of these filters or weights for the network, resulting in the identification of greater detail than is humanly possible.

  • Easier feature extraction. Convolutional layers within a CNN are good feature extractors adept at uncovering interesting patterns in any area of an image as they are represented by continuous blocks of pixels. Convolutional neural networks can learn image features from raw data extraction, eliminating the need for manual image processing.

  • More efficient computation. The concepts of parameter sharing (weight sharing of all neurons in a feature map) and local connectivity (neural networks connected to subsets of input images) are unique to convolutional neural networks. They make computation more efficient by helping to reduce the number of parameters defining a system.


Related content