en.wikipedia.org/wiki/VGGNet
1 correction found
a decision that was only revised in ConvNext (2022).
This is too absolute. CNN architectures had already moved away from an all-3×3 design years before ConvNeXt; for example, GoogLeNet/Inception (2014) used 1×1, 3×3, and 5×5 convolutions.
Full reasoning
The word "only" makes this claim incorrect.
Long before ConvNeXt was introduced in 2022, major CNN architectures had already revised VGG's all-3×3 pattern. The clearest counterexample is GoogLeNet / Inception from 2014, whose architecture explicitly combined 1×1, 3×3, and 5×5 convolutions. That means the design choice was not first revisited in ConvNeXt.
In other words:
- VGG (2014) emphasized stacking small 3×3 convolutions.
- Inception / GoogLeNet (2014) already departed from that by using mixed kernel sizes, including 5×5.
- ConvNeXt (2022) is a much later ConvNet redesign, not the first revision of the "just 3×3" approach.
So the article's statement is inaccurate because it dates the first revision of that kernel-size decision to 2022, even though a prominent counterexample exists from 2014.
3 sources
- Going Deeper with Convolutions
1x1 convolutions ... the expensive 3×3 and 5×5 convolutions. ... deeper and wider Inception network with slightly superior
- [1409.4842] Going Deeper with Convolutions
Submitted on 17 Sep 2014 ... We propose a deep convolutional neural network architecture codenamed "Inception" ... responsible for setting the new state of the art for classification and detection in the ImageNet Large-Scale Visual Recognition Challenge 2014 (ILSVRC 2014).
- [2201.03545] A ConvNet for the 2020s
Submitted on 10 Jan 2022 ... The outcome of this exploration is a family of pure ConvNet models dubbed ConvNeXt.