ResNeXt
ResNeXt: Aggregated Residual Transformations for Deep Neural Networks · Also known as: ResNeXt, Aggregated Residual Transformations, grouped convolution residual network, cardinality-based ResNet
ResNeXt is a deep convolutional neural network architecture introduced by Xie, Girshick, Dollár, Tu, and He at CVPR 2017. It extends the residual network (ResNet) design by introducing a new architectural dimension called cardinality — the number of independent, parallel transformation paths within each residual block — enabling higher accuracy with fewer parameters and a simpler, more uniform design than its predecessors.
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.
When to use it
ResNeXt is well-suited for image classification, object detection, instance segmentation, and other vision tasks where a strong pretrained convolutional backbone is required. It is an appropriate choice when ResNet accuracy has been exhausted and the practitioner wants a principled improvement without a dramatic increase in parameters or architectural complexity. The architecture assumes sufficient labelled image data (typically hundreds of thousands of examples for training from scratch, or any size when fine-tuning a pretrained checkpoint). It is less appropriate for tasks that require architectures with attention mechanisms (e.g., Vision Transformers) or for resource-severely constrained deployments where lightweight models such as MobileNet are necessary.
Strengths & limitations
- Cardinality as a new, effective scaling dimension: increasing the number of parallel paths consistently improves accuracy at equal parameter and FLOP budgets.
- Simpler and more uniform design than Inception: all parallel branches within a block are identical, eliminating the need to hand-tune heterogeneous branch configurations.
- Efficient implementation via grouped convolution, which maps directly to optimised GPU primitives.
- Strong pretrained checkpoints available for ImageNet, making fine-tuning straightforward.
- Generalises well across tasks: the ResNeXt backbone transfers effectively to detection (Faster R-CNN, Mask R-CNN) and segmentation pipelines.
- Like all large convolutional networks, ResNeXt requires substantial computational resources (GPU memory and training time) for training from scratch.
- The grouped convolution implementation can be sensitive to the choice of C and d; suboptimal cardinality values may not yield the expected gains.
- On tasks that benefit strongly from global attention (e.g., long-range spatial dependencies), Vision Transformer architectures may outperform ResNeXt.
- Lacks built-in channel attention (unlike SE-ResNeXt), which can limit performance on fine-grained classification without additional modules.
Frequently asked
What is cardinality, and why does it matter?
Cardinality C is the number of independent, parallel transformation paths (groups) within each residual block. Xie et al. (2017) showed empirically that increasing C while holding the total parameter count and FLOPs constant consistently improves accuracy more than equivalent increases in depth or width, establishing cardinality as a third effective dimension for scaling CNNs.
How is ResNeXt different from ResNet?
ResNet applies a single transformation per residual block. ResNeXt replaces that single transformation with an aggregated sum of C parallel, identical transformations implemented as a grouped convolution. The shortcut connection and overall block structure remain the same; the key change is inside the transformation path.
How does ResNeXt relate to Inception networks?
Both architectures use multiple parallel branches within a block. The key difference is homogeneity: Inception blocks use heterogeneous branches (different kernel sizes, pooling, etc.) that must be tuned independently, whereas ResNeXt uses identical branches that differ only in their channel group, making the design far simpler and easier to scale.
Which ResNeXt variant should I start with?
ResNeXt-50 (32×4d) is a practical starting point: it has a similar parameter count to ResNet-50 but higher accuracy, and pretrained weights are widely available. ResNeXt-101 (32×8d or 64×4d) offers further gains at greater computational cost and is commonly used as a strong detection backbone.
Sources
- Xie, S., Girshick, R., Dollár, P., Tu, Z., & He, K. (2017). Aggregated Residual Transformations for Deep Neural Networks. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 5987–5995. DOI: 10.1109/CVPR.2017.634 ↗
- 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 ↗
- Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep Learning. MIT Press. ISBN: 978-0-26-203561-3
How to cite this page
ScholarGate. (2026, June 3). ResNeXt: Aggregated Residual Transformations for Deep Neural Networks. ScholarGate. https://scholargate.app/en/deep-learning/resnext
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.
- DenseNetDeep learning↔ compare
- EfficientNetDeep learning↔ compare
- MobileNetDeep learning↔ compare
- ResNetDeep learning↔ compare