Convolutional Neural Network

A Convolutional Neural Network (CNN) is a Neural Network architecture widely used in Deep Learning to process grid-structured data, especially images. Its Convolutional Layers apply learnable Filters or Convolution Kernels across the input to extract local features, such as edges and textures, with successive layers learning increasingly complex patterns. Compared with a Multi-Layer Perceptron of comparable dimensions, Weight Sharing and Local Receptive Fields reduce the number of parameters and improve computational efficiency. These properties support Translation Equivariance—shifting the input shifts the resulting Feature Maps—rather than guaranteeing Translation Invariance, where outputs remain unchanged. Exact equivariance depends on implementation details and can be disrupted by Stride, Pooling, Padding, and boundary effects. Pooling, global aggregation, or learning from training data can encourage partial translation invariance, but standard CNNs are not inherently invariant. Typical architectures combine convolutional layers with Activation Functions such as ReLU, pooling layers for Downsampling, and often Fully Connected Layers for Classification or Regression. Pioneered by researchers including Yann LeCun through architectures such as LeNet, and subsequently advanced by AlexNet, VGGNet, and ResNet, CNNs became foundational to Computer Vision and remain widely used in systems developed by companies and research organizations such as NVIDIA, Google DeepMind, and Meta AI.