All corrections
en.wikipedia.org/wiki/VGGNet
1 correction found
1
Claim
a decision that was only revised in ConvNext (2022).
Correction
This is incorrect: major CNN architectures revised the all-3×3 pattern well before 2022, including Inception/GoogLeNet and ResNet.
Full reasoning
This sentence says the move to all-3×3 convolutions was “only revised in ConvNext (2022)”, but well-known CNNs changed that design years earlier.
Two clear counterexamples from primary sources:
- Inception / GoogLeNet (ILSVRC 2014) used multiple kernel sizes, including 1×1, 3×3, and 5×5 convolutions.
- ResNet (CVPR 2016) used a 7×7 convolution in its stem.
Because these architectures reintroduced non-3×3 kernels long before 2022, the claim that this design choice was only revised in ConvNeXt is factually wrong.
2 sources
- Going Deeper with Convolutions
1x1 convolutions ... the expensive 3×3 and 5×5 convolutions.
- Deep Residual Learning for Image Recognition
7x7 conv, 64, /2 ...