A Convolutional Neural Network (CNN) is a class of artificial neural networks (ANN) in the field of Deep Learning. They are most commonly used for visual image analysis, such as image classification, image segmentation, image and video recognition, and medical image analysis. They are also used in recommender systems, natural language processing, financial time series, or in combination with other types, especially long-term short-term memory (LSTM).

Convolutional networks were inspired by biological processes in that the connectivity pattern between neurons resembles the organization of the visual cortex of animals. Intuitively, CNNs are based on the architecture of convolution kernels or shared-weight filters that slide along input features and provide translation-equivariant responses known as feature maps. CNNs are often known as space invariant artificial neural networks (SIANN), however most convolutional neural networks are indeed only equivariant to translation. The network then learns to optimize the filters through automated learning, whereas in traditional algorithms these filters are hand. This independence from prior knowledge and human intervention in feature extraction is a major advantage.

CNNs are regularized versions of multilayer perceptrons. Multilayer perceptrons generally refer to fully connected networks, i.e., every neuron in one layer is connected to all neurons in the next layer, which makes them prone to data overfitting due to over-parameterization. Typical methods to regularize or prevent overfitting include penalizing parameters during training (e.g., weight decay) or trimming connectivity (skipped connections, dropout, etc.). CNNs take a different approach towards regularization, taking advantage of hierarchical patterns in the data and assembling patterns of increasing complexity from smaller and simpler patterns imprinted in their filters. Working only with patterns, CNNs are therefore at the lower extreme on a scale of connectivity and complexity. In the modern CNN architectures, otherwise, all of these regularizations can be found.