What Are Convolutional Neural Networks?

Convolutional neural networks are a class of deep learning models designed to process data with spatial or temporal structure, particularly two-dimensional images, by applying learnable linear filters, called convolution kernels, to the input and extracting hierarchical feature representations through successive layers. Each convolutional layer computes local dot products between the input and a set of learned kernels, then applies a nonlinear activation function to produce feature maps that encode the presence of patterns such as edges, textures, and eventually complex object parts at different spatial scales. The weights of these kernels are shared across all spatial positions, which dramatically reduces the number of parameters compared to a fully connected network and encodes the prior knowledge that useful visual features can appear anywhere in an image.

Convolutional neural networks draw on earlier work in biological vision research and the study of receptive fields in the mammalian visual cortex. Yann LeCun and collaborators demonstrated practical backpropagation training of convolutional networks for handwritten digit recognition in 1989, and the 2012 AlexNet architecture, which applied deep convolutional networks to the ImageNet large-scale recognition challenge, established CNNs as the dominant approach in computer vision and catalyzed the broader deep learning resurgence.

Convolutional Layer Architecture

The convolutional layer is the defining component of a CNN. A kernel is a small matrix, typically 3x3 or 5x5 elements, that slides across the input feature map with a fixed stride, computing at each position the dot product between the kernel weights and the input patch. Multiple kernels operate in parallel, producing multiple feature maps from a single input layer. Depth, stride, padding, and the number of kernels are hyperparameters that control the spatial resolution, receptive field size, and representational capacity of each layer. The landmark AlexNet paper on ImageNet classification introduced several architectural choices, including ReLU activations and dropout regularization, that became standard practice. Deeper networks such as VGGNet, ResNet, and EfficientNet improved accuracy by stacking more layers or introducing residual connections that allow gradients to flow through hundreds of layers during training.

Pooling and Fully Connected Layers

Pooling layers follow convolutional layers and reduce the spatial dimensions of feature maps by summarizing local neighborhoods, either by taking the maximum value (max pooling) or the average. This downsampling decreases computational cost in subsequent layers, reduces sensitivity to small translations of the input, and progressively increases the spatial extent of each unit's receptive field. After several alternating convolutional and pooling layers, the feature maps are flattened into a vector and passed through one or more fully connected layers that combine the extracted features to produce the final output, such as class probability scores. Batch normalization, introduced in 2015, is typically inserted between convolutional and activation layers to stabilize training by normalizing internal layer activations. The IEEE Xplore paper on CNN for image detection and recognition surveys architectural variants and their performance characteristics across benchmark datasets.

Training and Generalization

CNNs are trained by minimizing a loss function, such as categorical cross-entropy for classification, using backpropagation to compute gradients and stochastic gradient descent or an adaptive variant such as Adam to update weights. Transfer learning, in which a network pretrained on a large dataset such as ImageNet is fine-tuned on a smaller domain-specific dataset, allows effective training when labeled data are limited. CNNs are also central to generative adversarial networks, where a convolutional generator produces synthetic images and a convolutional discriminator distinguishes real from generated examples. IBM's overview of convolutional neural networks discusses the range of tasks addressed by CNNs beyond image classification, including object detection, segmentation, and video understanding.

Applications

Convolutional neural networks have applications across many domains, including:

• Computer vision, including image classification, object detection, and semantic segmentation for autonomous vehicles and surveillance systems
• Medical imaging, where CNNs detect lesions, classify tissue types, and assist radiologists in reading CT, MRI, and pathology slides
• Natural language processing, where one-dimensional convolutional filters applied to word embedding sequences perform text classification and sentiment analysis
• Speech and audio processing, where convolutional front ends extract spectral features from mel-spectrogram representations of audio signals