What is a convolutional neural network (CNN)?

A convolutional neural network (CNN) is a type of deep learning model designed for processing grid-like data, such as images. Unlike traditional neural networks that treat input data as flat vectors, CNNs preserve the spatial structure of data by using convolutional layers. These layers apply filters (small grids of weights) to local regions of the input, detecting patterns like edges, textures, or shapes. For example, a filter might scan an image to identify vertical edges in the first layer. This approach mimics how the human visual system processes information hierarchically, starting with simple features and combining them into complex structures.

CNNs are built with three core components: convolutional layers, pooling layers, and fully connected layers. Convolutional layers perform feature extraction by sliding filters across the input and computing dot products to produce feature maps. For instance, a 3x3 filter applied to a 32x32 pixel image generates a feature map highlighting where specific patterns occur. Pooling layers (like max-pooling) then downsample these maps, reducing dimensionality while preserving important features—this helps control computational costs and prevents overfitting. Finally, fully connected layers interpret the extracted features for tasks like classification. For example, after several convolutional and pooling layers, a CNN might flatten the output and use dense layers to classify an image as “cat” or “dog.”

CNNs excel in computer vision tasks because they efficiently handle spatial hierarchies and translation invariance. Translation invariance means a cat detected in the top-left corner of an image is recognized as a cat even if it later appears in the bottom-right. This is achieved through weight sharing in convolutional filters, which drastically reduces parameters compared to fully connected networks. Practical applications include image recognition (e.g., ResNet), object detection (e.g., YOLO), and medical imaging (e.g., tumor detection). Frameworks like TensorFlow and PyTorch simplify CNN implementation by providing prebuilt layers and optimization tools. For developers, understanding hyperparameters like filter size, stride, and padding is key to designing effective models.