A convolutional neural network, or CNN, is a Neural network architecture built around convolutional layers that apply small, shared filters across an input, making it especially effective for grid-structured data such as images, where nearby pixels are related and useful patterns, like edges or textures, can appear anywhere in the frame. By reusing the same filter weights across every spatial location, CNNs need far fewer parameters than a fully connected network of comparable size and naturally encode the assumption that an image's meaning is largely unchanged by translating its content, a property known as translation invariance.
History
The architecture's conceptual roots go back to Kunihiko Fukushima's neocognitron, a 1980 hierarchical model of the visual cortex, described by Kunihiko Fukushima, that introduced layered feature detection without yet using Backpropagation for training. Yann LeCun combined convolutional layers with backpropagation in the late 1980s, producing LeNet, which by the mid-1990s could read handwritten digits on bank checks in commercial deployment, one of the earliest practical successes of Deep learning. CNNs remained a specialized tool for roughly two decades until 2012, when AlexNet, designed by Alex Krizhevsky with Ilya Sutskever and Geoffrey Hinton, won the ImageNet 2012 (AlexNet moment) ImageNet competition by a wide margin using a deep CNN trained on GPU (in AI)s, an event widely credited with igniting the modern deep learning boom.
Architecture
A typical CNN stacks alternating convolutional layers, which detect local patterns using learned filters, and pooling layers, which downsample the spatial resolution while retaining the strongest detected features, progressively building from low-level patterns like edges and colors in early layers to high-level, semantic concepts like faces or objects in deeper layers. The final layers are usually fully connected, mapping the extracted features to output classes. Later architectures such as VGGNet, ResNet, and Inception pushed depth much further, with ResNet's 2015 introduction of residual, or skip, connections solving the vanishing-gradient problem that had previously made very deep CNNs difficult to train with ordinary Gradient descent.
Applications
CNNs became the default architecture for Computer vision tasks including image classification, object detection, semantic segmentation, and facial recognition, and they underpin components of self-driving perception systems, medical image analysis, and early Generative adversarial network designs, where CNNs typically formed both the generator and discriminator. They have also been applied outside vision, for instance to audio spectrograms and some text classification tasks, wherever data has a local, grid-like structure a shared filter can exploit.
Decline in relative dominance
Since around 2020, Vision Transformers, which adapt the Transformer (architecture) architecture and its attention mechanism to image patches instead of convolutional filters, have matched or exceeded CNN performance on many large-scale vision benchmarks, particularly when trained on very large datasets, prompting debate about whether convolution's built-in spatial assumptions remain necessary once enough data and compute are available. CNNs nonetheless remain widely used in practice, especially in resource-constrained or Edge AI settings, because their inductive biases make them more data-efficient than transformers when training data is limited, and they continue to serve as encoders inside larger multimodal systems.