Research map / Architectures and efficiency / CNN design
InceptionNeXt: When Inception Meets ConvNeXt
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes InceptionNeXt, which replaces ConvNeXt's 7x7 depthwise convolution with an Inception-style depthwise convolution (3x3, two orthogonal band kernels, identity branch) to gain speed while keeping ImageNet accuracy.
Problem
Large-kernel depthwise convolutions (e.g., ConvNeXt's 7x7) have few FLOPs but high memory-access cost, making CNNs slow on GPUs; shrinking the kernel speeds them up but hurts accuracy.
Approach
Split channels into four groups processed in parallel: a small 3x3 depthwise conv, 1x11 and 11x1 band depthwise convs, and an identity branch (only 1/8 of channels are convolved), yielding a large receptive field at linear cost in kernel size and channels.
Key results
- InceptionNeXt-T reaches 82.3% top-1 on ImageNet-1K, +0.2% over ConvNeXt-T (82.1%), with 1.6x training and 1.2x inference throughput on A100.
- InceptionNeXt-B achieves 84.0% top-1 on ImageNet-1K (+0.2% over ConvNeXt-B) with higher throughput.
- InceptionNeXt-T with UperNet reaches 47.9 mIoU on ADE20K vs 46.7 for ConvNeXt-T.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace ConvNeXt's 7×7 depthwise convolution with Inception depthwise convolution | ConvNeXt-T · ImageNet-1K val | top-1 accuracy: 82.1 -> 82.3 | Helped |
| replace ConvNeXt's 7×7 depthwise convolution with Inception depthwise convolution | ConvNeXt-T · ImageNet-1K | A100 training throughput: 575 -> 901 | Helped |
| use InceptionNeXt-T instead of ConvNeXt-T as the backbone | UperNet · ADE20K val | mIoU: 46.7 -> 47.9 | Helped |
| use InceptionNeXt-T instead of PoolFormer-S24 as the backbone | Semantic FPN · ADE20K val | mIoU: 40.3 -> 43.1 | Helped |
| add the horizontal band-kernel branch (ablation) | InceptionNeXt-T · ImageNet-1K val | top-1 accuracy: 81.9 -> 82.3 | Helped |
| add the vertical band-kernel branch (ablation) | InceptionNeXt-T · ImageNet-1K val | top-1 accuracy: 81.9 -> 82.3 | Helped |
| add the small square-kernel branch (ablation) | InceptionNeXt-T · ImageNet-1K val | top-1 accuracy: 82.0 -> 82.3 | Helped |
| place horizontal and vertical band kernels in sequence instead of parallel (ablation) | InceptionNeXt-T · ImageNet-1K val | top-1 accuracy: 82.3 -> 82.1 | Hurt |
| increase band kernel size from 7 to 11 (ablation) | InceptionNeXt-T · ImageNet-1K val | top-1 accuracy: 82.1 -> 82.3 | Helped |
| increase band kernel size from 11 to 13 (ablation) | InceptionNeXt-T · ImageNet-1K val | top-1 accuracy: 82.3 -> 82.0 | Hurt |
| increase convolution branch ratio from 1/8 to 1/4 (ablation) | InceptionNeXt-T · ImageNet-1K val | top-1 accuracy: 82.3 -> 82.2 | No clear effect |
| decrease convolution branch ratio from 1/8 to 1/16 (ablation) | InceptionNeXt-T · ImageNet-1K val | top-1 accuracy: 82.3 -> 81.8 | Hurt |
Surprising findings
- Removing the 3x3 square-kernel branch only drops top-1 to 82.0%, suggesting the band kernels carry most of the benefit.
- MetaNeXt with self-attention as token mixer failed to converge (3.9% accuracy), showing the simplified block tolerates only light token mixers.
Limitations
- Speed gains shrink as model size grows because MLPs dominate computation at larger channel counts.
- Parallel band-kernel branches are implemented at a high level in PyTorch and sequential form was slightly faster, so the speedups depend on implementation maturity.
- Larger band kernel (13) decreases accuracy, which the authors attribute to optimization issues rather than fully explaining.
Experimental setup
Datasets: ImageNet-1K ADE20K
Benchmarks: ImageNet-1K ADE20K
Metrics: top-1 accuracy training throughput inference throughput mIoU FPS MACs parameters
Baselines: ConvNeXt Swin Transformer ResNet-50 RSB-ResNet RegNetY FocalNet RepLKNet PoolFormer PVT DeiT
Builds on: ConvNeXt MetaFormer Inception v3 ShuffleNet V2 DeiT
Closest papers by meaning
- More ConvNets in the 2020s: Scaling up Kernels Beyond 51x51 using Sparsity (ICLR 2023 · similarity 0.88)
- Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning (AAAI 2017 · similarity 0.88)
- Scaling Up Your Kernels to 31x31: Revisiting Large Kernel Design in CNNs (CVPR 2022 · similarity 0.87)
- Boosting the Throughput and Accelerator Utilization of Specialized CNN Inference Beyond Increasing Batch Size (ICML 2021 · similarity 0.87)
- UniRepLKNet: A Universal Perception Large-Kernel ConvNet for Audio Video Point Cloud Time-Series and Image Recognition (CVPR 2024 · similarity 0.86)
- PeLK: Parameter-efficient Large Kernel ConvNets with Peripheral Convolution (CVPR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.