Convolutional Neural Networks

A convolutional neural network (CNN or ConvNet) is a deep learning architecture built for data laid out on a grid and, above all, for images. Where a fully connected network links every neuron to all the units in the previous layer, a CNN assumes that neighbouring pixels are related and exploits that locality to recognise visual patterns. That simple but powerful assumption is what turned CNNs into the tool that unlocked modern computer vision.

Their core idea is to learn a useful representation straight from the pixels, without an engineer hand-crafting the features. The network figures out what to look at on its own, layer by layer, tuning those detectors as it trains.

Convolution: local filters and shared weights

The building block that gives the architecture its name is the convolution. A filter, or kernel, is a small matrix of weights —say 3×3— that slides across the whole image and, at each position, computes a combination of the pixels it covers. The output is an activation map showing where the pattern the filter responds to appears: an edge, a blob of colour, a texture. Each layer applies many filters in parallel, so a single input yields dozens of complementary maps.

Two properties explain their power. The first is local connectivity: each neuron sees only a small region, its receptive field, rather than the entire image. The second is weight sharing: the same filter is reused at every position. Sharing weights slashes the parameter count compared with a dense layer and grants translation invariance —an eye or a wheel is detected whether it sits at the top or the bottom of the scene.

Pooling and the hierarchy of features

Between convolutions sit pooling, or subsampling, layers that summarise each region by taking, for instance, its maximum value. This lowers the map's resolution, cuts computation and makes the representation more robust to small shifts. By stacking these blocks, the network builds a hierarchy of features: early layers respond to edges and colours; middle layers to motifs and parts —a corner, an eye, a wheel—; and deep layers to whole objects. That progression from simple to complex is what carries the model from pixels to concepts, and what echoes the early stages of the biological visual system.

History: from LeNet to AlexNet (2012) and today

The inspiration traces back to Hubel and Wiesel's work on the visual cortex, which described neurons tuned to oriented edges within local receptive fields. Building on that idea, Yann LeCun developed LeNet-5 (1998), a CNN trained with backpropagation that read handwritten digits on cheques and postal codes. The turning point came in 2012, when AlexNet, by Krizhevsky, Sutskever and Hinton, won the ImageNet competition, cutting the classification error well below its rivals' and proving the power of training deep networks on GPUs. Deeper architectures such as VGG and ResNet followed. Today CNNs share the stage with Vision Transformers, which bring attention mechanisms to images, yet they remain an essential landmark of computer vision.