What is a convolutional neural network?

A convolutional neural network (CNN) is a type of deep learning model designed to process grid-structured data, such as images, videos, or audio spectrograms. Unlike traditional neural networks that treat input data as flat vectors, CNNs preserve spatial relationships by applying convolutional operations. These operations use small filters (or kernels) that slide across the input to detect local patterns like edges, textures, or shapes. For example, a filter might activate when it encounters a vertical edge in an image. By stacking multiple convolutional layers, CNNs hierarchically learn increasingly complex features, from simple lines to object parts and entire objects.

CNNs are built with three core components: convolutional layers, pooling layers, and activation functions. Convolutional layers apply filters to extract features, while pooling layers (like max-pooling) reduce spatial dimensions, making the model computationally efficient and robust to small shifts in input. Activation functions such as ReLU (Rectified Linear Unit) introduce non-linearity, enabling the network to learn complex relationships. For instance, a typical CNN for image classification might start with a convolutional layer detecting edges, followed by a pooling layer to downsample the feature maps, then another convolutional layer to combine edges into shapes like circles or corners. Finally, fully connected layers use these high-level features for tasks like classification.

CNNs excel in computer vision tasks such as image recognition, object detection, and segmentation. For example, a self-driving car system might use a CNN to identify pedestrians in camera feeds by analyzing pixel patterns. Their efficiency comes from parameter sharing (the same filter scans the entire input) and spatial hierarchy (progressively abstracting details). Beyond images, CNNs adapt to other domains: 1D convolutions process time-series data, and 3D convolutions analyze video frames. Frameworks like TensorFlow or PyTorch simplify CNN implementation by providing pre-built layers and optimization tools. While initially designed for vision, CNNs remain versatile due to their ability to model local dependencies in any grid-like data.