en.wikipedia.org/wiki/VGGNet
1 correction found
a decision that was only revised in ConvNext (2022).
This is incorrect because CNN architectures returned to larger kernels well before 2022. GoogLeNet/Inception (2015, used for ILSVRC 2014) used 5×5 and 7×7 convolutions, and ResNet (2016) also begins with a 7×7 convolution.
Full reasoning
VGG did popularize extensive use of 3×3 convolutions, but the claim that this design choice was only revised in ConvNeXt in 2022 is contradicted by earlier primary sources.
- The official VGG project page says the VGG models use very small 3×3 filters in all convolutional layers.
- The official CVPR 2015 paper for GoogLeNet/Inception shows the architecture already using 5×5 and 7×7 convolutions.
- The official CVPR 2016 ResNet paper shows the ImageNet architecture starting with a 7×7 convolution.
So larger-than-3×3 kernels had already reappeared in major CNN architectures years before 2022. That makes the statement that this decision was "only revised in ConvNext (2022)" factually wrong.
3 sources
- Visual Geometry Group - University of Oxford: Very Deep Convolutional Networks for Large-Scale Visual Recognition
To reduce the number of parameters in such very deep networks, we use very small 3×3 filters in all convolutional layers (the convolution stride is set to 1).
- Going Deeper With Convolutions (CVPR 2015)
Figure 3: GoogLeNet network ... Conv 3x3+1(S) ... Conv 5x5+1(S) ... 7x7+1(V).
- Deep Residual Learning for Image Recognition (CVPR 2016)
Example network architectures for ImageNet ... output size: 112 ... 7x7 conv, 64, /2.