ResNet-18

ResNet-18 is a streamlined convolutional neural network that uses residual blocks to enable efficient deep learning for image classification.

Microsoft Research developed ResNet-18 to address the vanishing gradient problem using skip connections. This architecture features 18 weighted layers (17 convolutional and one fully connected) and handles roughly 1.8 billion FLOPs per pass. It remains a primary baseline for computer vision because it balances 11.7 million parameters with high accuracy on the ImageNet dataset. Engineers often select this model for real-time applications or as a starting point for transfer learning due to its low latency and reliable performance.

https://arxiv.org/abs/1512.03385