ResNet (Residual Network)
Residual Network (ResNet) · Also known as: ResNet, Residual Network, Deep Residual Learning, ResNet-50, ResNet-101, ResNet-152, He-net
ResNet (Residual Network) is a deep convolutional neural network architecture introduced by Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun at CVPR 2016. By inserting shortcut (skip) connections that carry the input of a block directly to its output — defining the block's task as learning a residual correction rather than a full mapping — ResNet enabled training of networks with hundreds or even thousands of layers without the vanishing-gradient degradation that had previously made very deep networks impractical. It won the ILSVRC 2015 image recognition competition with a top-5 error of 3.57% and remains the most widely used backbone architecture in computer vision.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
+1 more
When to use it
ResNet is appropriate whenever a task requires learning powerful visual representations from image or video data — particularly when high accuracy matters and a pre-trained backbone is acceptable. Standard use cases include image classification, object detection (as a backbone inside detectors such as Faster R-CNN or YOLO), semantic segmentation, and feature extraction for transfer learning. ResNet variants (ResNet-18/34 for speed; ResNet-50/101/152 for accuracy) can be chosen to match the compute budget. Pre-trained ImageNet weights are publicly available and enable effective transfer learning even with small task-specific datasets (a few hundred labeled images can suffice with fine-tuning). The main assumption is that the task is image-based; for tabular, sequential, or text data, other architectures are typically preferable.
Strengths & limitations
- Skip connections solve the vanishing-gradient problem, enabling stable training of networks with 50–1000+ layers.
- State-of-the-art performance on a wide range of computer vision benchmarks; won ILSVRC 2015 with 3.57% top-5 error.
- Pre-trained weights (ImageNet) allow highly effective transfer learning even with small task-specific datasets.
- Modular and scalable: the same block structure spans ResNet-18 to ResNet-1202, offering a compute-accuracy trade-off.
- Widely supported in every major deep-learning framework (PyTorch, TensorFlow/Keras, JAX), with production-ready implementations.
- Computationally expensive relative to lightweight architectures (MobileNet, EfficientNet-B0) — unsuitable for real-time deployment on resource-constrained hardware without further compression.
- The standard architecture was designed for fixed-size images; adapting to variable-resolution or non-image inputs requires additional engineering.
- Pre-trained weights are biased toward ImageNet categories; transfer to very different domains (e.g., medical images, remote sensing) may require extensive fine-tuning or domain-specific pre-training.
- Despite skip connections, extremely deep variants (ResNet-1202) can still suffer from optimisation difficulties and may underperform shallower counterparts.
Frequently asked
Which ResNet variant should I choose?
ResNet-18 and ResNet-34 use the two-layer basic block and are fast enough for real-time or resource-limited settings. ResNet-50 is the most common trade-off — strong accuracy with reasonable compute. ResNet-101 and ResNet-152 are used when accuracy is the primary concern and compute is plentiful. For mobile or edge deployment, consider lighter architectures (MobileNetV3, EfficientNet-B0) or ResNet with pruning.
Do I need to train ResNet from scratch?
Rarely. Pre-trained ImageNet weights are publicly available for all standard ResNet variants and provide an excellent starting point via transfer learning. Fine-tuning the final layers (or the entire network at a small learning rate) typically outperforms training from scratch unless the target domain is radically different from ImageNet and a very large task-specific dataset is available.
Why does ResNet outperform plain deep networks of the same depth?
He et al. (2016) showed that plain networks suffer from a training-error degradation when depth increases — more layers produce higher training error, not just higher test error, so overfitting is not the cause. Skip connections allow the network to learn residuals rather than full mappings and provide a gradient highway that prevents the vanishing-gradient collapse, enabling deeper networks to actually benefit from their extra capacity.
Is ResNet still competitive with modern architectures?
ResNet-50 and ResNet-101 remain strong baselines and are widely used as backbones in detection and segmentation pipelines. Newer architectures (EfficientNet, Vision Transformer, ConvNeXt) achieve better accuracy-efficiency trade-offs on pure classification, but ResNet is often preferred for its simplicity, stability, and the large body of proven downstream pipelines built around it.
Sources
- He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770–778. DOI: 10.1109/CVPR.2016.90 ↗
- He, K., Zhang, X., Ren, S., & Sun, J. (2015). Deep Residual Learning for Image Recognition. arXiv:1512.03385. link ↗
- Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning (Ch. 9: Convolutional Networks). MIT Press. ISBN: 978-0-262-03561-3
How to cite this page
ScholarGate. (2026, June 3). Residual Network (ResNet). ScholarGate. https://scholargate.app/en/deep-learning/resnet
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- AlexNetDeep learning↔ compare
- DenseNetDeep learning↔ compare
- EfficientNetDeep learning↔ compare
- Inception NetworkDeep learning↔ compare