Convolutional Neural Networks
Networks that learn from images and grids using shared-weight filters: convolution, pooling, and the architectures that won computer vision.
A convolutional neural network (CNN) is a neural network specialized for grid-structured data — images above all, but also audio spectrograms, video, and volumetric medical scans. Its defining layer, the convolutional layer, replaces the dense connections of ordinary networks with small learned filters that slide across the input. A 3×3 filter applied at every position detects the same pattern everywhere: this weight sharing is what makes CNNs practical, because an image of 224×224×3 pixels would require millions of weights per neuron if processed densely.
Each filter learns to detect a feature, and stacking layers builds a hierarchy: early filters respond to edges, color blobs, and oriented lines; middle layers combine them into textures and parts; deep layers detect whole objects such as faces or wheels. The receptive field — the region of the input that influences a neuron — grows with depth, which is how a network with small filters acquires global awareness. Multiple filters per layer produce multiple feature maps, and pooling layers (traditionally max-pooling over 2×2 windows) downsample the maps, reducing computation and giving small translation invariance; modern architectures often replace pooling with strided convolutions.
CNNs are trained end-to-end with backpropagation and gradient descent — no hand-designed features. Their modern history is a sequence of milestones: LeNet-5 (Yann LeCun, 1998) read bank check digits; AlexNet (2012) cut the ImageNet error rate roughly in half and ignited the deep learning era; VGG showed the power of simple, deep stacks of 3×3 filters; and ResNet (2015) introduced residual connections that allowed hundreds of layers to train, still the backbone of many modern designs. Later lines such as EfficientNet and MobileNet traded accuracy against compute for mobile deployment.
CNNs dominate computer vision practice: image classification, object detection (R-CNN and its descendants, YOLO), semantic segmentation, face recognition, and super-resolution, plus applications in medical imaging, autonomous driving, and satellite analysis. Two practical techniques multiply their power: data augmentation (random crops, flips, color jitter) makes them robust, and transfer learning — starting from a network pretrained on ImageNet and fine-tuning on a smaller dataset — is the default way to apply them with limited data.
Their limitations are equally well documented: CNNs are sensitive to distribution shift and adversarial perturbations, require large labeled datasets, and process the whole image uniformly rather than with the focused attention humans use. The attention mechanism that addressed some of these limits — first added to CNNs and then replacing them — is the basis of the transformer architecture, which in its vision variant (ViT) has become the main rival to convolutional networks.
Tags
computer vision deep learning neural networks