ResNet18

ResNet18 is a residual neural network architecture introduced as part of the ResNet family in 2015 by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. It demonstrates that relatively shallow networks can benefit from residual learning, enabling effective training of deeper models by mitigating vanishing gradients through skip connections.

The network consists of 18 layers and uses a basic residual block. Each block contains two consecutive 3x3 convolutional layers with batch normalization and a ReLU activation after each convolution. A identity skip connection adds the block’s input to its output. If the input and output dimensions differ, the skip path uses a 1x1 convolution to match dimensions. The architecture begins with a 7x7 convolution with 64 filters and stride 2, followed by a 3x3 max pooling layer. This is followed by four stages containing two residual blocks each, with channel sizes 64, 128, 256, and 512. Downsampling occurs at the first block of each stage except the first. The network ends with global average pooling and a fully connected layer for 1000-class classification.

Parameter count for ResNet18 is about 11.7 million. It is commonly trained on ImageNet and widely used as a backbone for transfer learning in computer vision tasks. The design emphasizes simple, uniform residual blocks to enable efficient gradient flow and faster convergence compared with plain deep networks. ResNet18 remains a popular baseline due to its balance of accuracy and computational efficiency, and is often contrasted with deeper ResNet variants that use bottleneck blocks.