Convolutional Neural Network
A Convolutional Neural Network (CNN) is a deep learning architecture that uses learnable convolutional filters to automatically extract spatial features from grid-like data, most commonly images, making it the foundational architecture for…
Definition
A Convolutional Neural Network (CNN) is a deep learning architecture that uses learnable convolutional filters to automatically extract spatial features from grid-like data, most commonly images, making it the foundational architecture for computer vision.
Overview
CNNs are designed around the idea that useful patterns in images — edges, textures, shapes — are local and translation-invariant, meaning a pattern that matters in one part of an image is likely to matter the same way elsewhere. A convolutional layer applies small learnable filters (kernels) that slide across the input, computing dot products at each position to produce feature maps, allowing the same filter to detect a pattern regardless of where it appears in the image. This dramatically reduces the number of parameters compared to a fully connected layer processing the same input, since each filter's weights are shared across all spatial positions. A typical CNN stacks multiple convolutional layers, often interleaved with nonlinear activation functions (commonly ReLU) and pooling layers (such as max pooling), which downsample feature maps to reduce spatial dimensions and add a degree of translation invariance. Early layers tend to learn simple features like edges and color gradients, while deeper layers combine these into increasingly abstract, complex representations — textures, object parts, and eventually whole objects — before final fully connected or global pooling layers produce a classification or other task-specific output. Landmark CNN architectures trace the field's progress: LeNet-5 pioneered the approach for digit recognition in the 1990s, AlexNet's 2012 ImageNet win demonstrated deep CNNs' power at scale and helped ignite the deep learning boom, and subsequent architectures like VGGNet, ResNet (which introduced residual/skip connections enabling much deeper networks), and EfficientNet pushed both accuracy and efficiency further. CNNs remain the dominant architecture for many computer vision tasks — image classification, object detection, semantic segmentation, and medical image analysis — though Vision Transformers (ViTs), which apply the Transformer architecture directly to image patches, have become competitive or superior on large-scale datasets, and many modern vision systems now use hybrid or Transformer-based designs instead of, or alongside, pure CNNs.
Key Concepts
- Convolutional layers apply shared learnable filters to detect local, translation-invariant patterns
- Parameter sharing dramatically reduces model size versus fully connected layers
- Pooling layers (e.g. max pooling) downsample feature maps and add spatial robustness
- Hierarchical feature learning — edges and textures in early layers, complex shapes in deeper layers
- Landmark architectures: LeNet-5, AlexNet, VGGNet, ResNet, EfficientNet
- Residual/skip connections (ResNet) enable training of much deeper networks
- Well suited to grid-structured data: images, video frames, and spectrograms
- Foundation for object detection, segmentation, and image classification pipelines