VGGNet

VGGNet demonstrated that network depth is a critical factor for performance in image recognition. Its key insight was radically simple: replace large convolutional filters with stacks of small 3x3 filters, allowing much deeper networks without increasing computational cost per layer.

We use Smaller Filter and a Deeper Network

Before VGG, architectures like AlexNet used large filters (11x11, 5x5) to capture spatial patterns. VGG asked: what if we use only 3x3 filters and just stack more layers?

Approach	Filter size	Depth	Receptive field
AlexNet-style	11x11, 5x5	Shallow (5 conv layers)	Large per layer
VGG-style	3x3 only	Deep (16-19 conv layers)	Built up gradually

Why 3x3 Works

Two stacked 3x3 conv layers have the same receptive field as one 5x5 layer. Three stacked 3x3 layers match a 7x7 layer. But the stacked version is better:

One 7x7 filter:     7 x 7 x C = 49C parameters
Three 3x3 filters:  3 x (3 x 3 x C) = 27C parameters

Benefit	Why
Fewer parameters	27C vs 49C for the same receptive field
More non-linearity	Each layer has its own ReLU, so three decision boundaries instead of one
Better feature learning	Forces the network to build features compositionally

Architecture: VGG-16

VGG-16 has 16 weight layers (13 convolutional + 3 fully connected). The architecture follows a strict, repeating pattern:

Input: 224x224x3 (RGB image)

Block 1:  [3x3 conv, 64]  x 2  +  MaxPool
Block 2:  [3x3 conv, 128] x 2  +  MaxPool
Block 3:  [3x3 conv, 256] x 3  +  MaxPool
Block 4:  [3x3 conv, 512] x 3  +  MaxPool
Block 5:  [3x3 conv, 512] x 3  +  MaxPool

Flatten

FC: 4096
FC: 4096
FC: 1000 (ImageNet classes)

Softmax

Spatial Dimensions Through the Network

Block	Output size	Channels
Input	224 x 224	3
Block 1 + Pool	112 x 112	64
Block 2 + Pool	56 x 56	128
Block 3 + Pool	28 x 28	256
Block 4 + Pool	14 x 14	512
Block 5 + Pool	7 x 7	512
Flatten	25088	-
FC layers	4096	-
Output	1000	-

The pattern is consistent: spatial dimensions halve, channels double at each block. This became a standard design principle for later architectures.

VGG Variants

Model	Conv layers	FC layers	Total weight layers	Parameters
VGG-11	8	3	11	133M
VGG-13	10	3	13	133M
VGG-16	13	3	16	138M
VGG-19	16	3	19	144M

The vast majority of parameters (about 124M of 138M in VGG-16) live in the fully connected layers, not the convolutional layers. This later motivated architectures like GoogLeNet and ResNet to eliminate or shrink FC layers.

Why VGG Matters

VGG established principles that influenced nearly every architecture that followed:

Principle	Impact
Depth improves performance	Directly motivated ResNet's push to 100+ layers
Small filters are sufficient	3x3 became the standard conv filter size
Uniform architecture	Simple repeating blocks became a design pattern
Halve spatial, double channels	Adopted by ResNet, Swin, and many others
Pretrained feature extractor	VGG features are still used today for perceptual loss, style transfer, and transfer learning

Limitations

Limitation	Detail
Massive parameter count	138M parameters, most wasted in FC layers
No skip connections	Cannot train much deeper without vanishing gradients
Slow inference	Heavy computation compared to later efficient architectures
No batch normalisation	Predates BatchNorm, making training harder and slower

VGG vs. Later Architectures

	VGG-16	ResNet-50	EfficientNet-B0
Parameters	138M	25M	5.3M
Layers	16	50	Compound scaled
Skip connections	No	Yes	Yes
Top-5 accuracy (ImageNet)	92.7%	93.3%	93.3%
Key innovation	Depth with 3x3 filters	Residual connections	Architecture search

ResNet achieves better accuracy with 5x fewer parameters and 3x more layers, precisely because skip connections solved the training difficulties VGG could not overcome.

JJ

Explorer