Yann LeCun et al. published Gradient-based learning applied to document recognition, introducing LeNet-5—a multi-layer convolutional network trained end-to-end with Backpropagation that became a blueprint for modern visual recognition.
Convolutional Neural Networks
Also known as: CNN · ConvNet · AlexNet · ResNet
Plain English
A neural network architecture designed for images and spatial grids. Instead of connecting every pixel to every neuron, it slides small pattern-detecting filters across the image to recognize edges, textures, and objects.
Technical Overview
Replaces dense all-to-all layers with weight-shared spatial convolution kernels and downsampling, providing translation equivariance and dramatic parameter efficiency on high-resolution grids. Deep CNNs like AlexNet and ResNet dominated computer vision for a decade and remain core components inside VAE encoders/decoders and U-Net Diffusion Models alongside Vision Transformers.
Timeline (4)
1998
Paper2012
PaperAlex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton published ImageNet Classification with Deep Convolutional Neural Networks (AlexNet), combining ReLU activations, dropout, and two NVIDIA GTX 580 GPUs to slash ImageNet top-5 error to 15.3%.
2015
PaperKaiming He et al. introduced ResNet (Deep Residual Learning for Image Recognition), adding identity skip connections so gradients could propagate cleanly through 100+ layers without vanishing—a residual block pattern later adopted by Transformers.
2022
PaperZhuang Liu et al. at Meta AI and UC Berkeley published A ConvNet for the 2020s (ConvNeXt), modernizing standard ResNets with Vision Transformer design choices (
7x7depthwise convolutions and inverted bottlenecks) to match hierarchical ViTs on detection and segmentation.
See also (9)
Open full graphConvolutional Neural Networks