Research map / Architectures and efficiency / CNN design
Rethinking the Inception Architecture for Computer Vision
Check your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use RMSProp training (ablation) | Inception-v2 · ILSVRC 2012 validation, single crop | top-1 error: 23.4% -> 23.1% | Helped |
| Add label-smoothing regularization (ablation) | Inception-v2 with RMSProp · ILSVRC 2012 validation, single crop | top-1 error: 23.1% -> 22.8% | Helped |
| Add label-smoothing regularization (ablation) | Inception-v2 with RMSProp · ILSVRC 2012 validation, single crop | top-5 error: 6.3 -> 6.1 | Helped |
| Factorize the first 7×7 convolution into a sequence of 3×3 convolutions (ablation) | Inception-v2 with label smoothing · ILSVRC 2012 validation, single crop | top-1 error: 22.8% -> 21.6% | Helped |
| Batch-normalize the auxiliary classifier fully connected layer (ablation) | Inception-v2 with factorized 7×7 · ILSVRC 2012 validation, single crop | top-1 error: 21.6% -> 21.2% | Helped |
| Use ReLU rather than a linear activation in the first factorization layer (ablation) | Inception factorization control · ImageNet validation | top-1 accuracy: 76.2% -> 77.2% | Helped |
| Batch-normalize the auxiliary classifier side head (ablation) | Inception auxiliary classifier · ImageNet validation | top-1 accuracy: +0.4% absolute | Helped |
| Remove the lower auxiliary branch (ablation) | Inception with two auxiliary classifiers · ImageNet classification | final quality | No clear effect |
| Use a smaller 79×79 receptive field instead of 299×299 at approximately constant computation (ablation) | Inception low-resolution input · ILSVRC 2012 validation | top-1 accuracy: 76.6% -> 75.2% | Hurt |
| Use a smaller 151×151 receptive field instead of 299×299 at approximately constant computation (ablation) | Inception low-resolution input · ILSVRC 2012 validation | top-1 accuracy: 76.6% -> 76.4% | Hurt |
| Use 144-crop evaluation rather than single-model 12-crop evaluation (ablation) | Inception-v3 · ILSVRC 2012 validation | top-1 error: 4.48% -> 4.2% | Helped |
| Ensemble four models and use multi-crop evaluation (ablation) | Inception-v3 · ILSVRC 2012 validation | top-5 error: 4.2% -> 3.46% | Helped |
Closest papers by meaning
- Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning (AAAI 2017 · similarity 0.92)
- Ubernet: Training a Universal Convolutional Neural Network for Low-, Mid-, and High-Level Vision Using Diverse Datasets and Limited Memory (CVPR 2017 · similarity 0.89)
- Involution: Inverting the Inherence of Convolution for Visual Recognition (CVPR 2021 · similarity 0.89)
- EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks (ICML 2019 · similarity 0.88)
- KernelWarehouse: Rethinking the Design of Dynamic Convolution (ICML 2024 · similarity 0.87)
- LSNet: See Large, Focus Small (CVPR 2025 · similarity 0.87)
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.