CS5720 - Week 5
Slide 84 of 100

VGGNet: Going Deeper with Small Filters

The VGG Philosophy

"What if we only used 3×3 convolutions throughout the entire network?"

This simple yet powerful idea led to VGGNet's elegant and effective architecture.
Key Design Principles:
• Use only 3×3 convolutional filters
• Stack many layers to go deeper
• Double filters after each pooling
• Keep architecture extremely uniform
• Simple is better than complex
The Result:
2nd place in ILSVRC 2014 classification, but became more influential than the winner due to its simplicity and effectiveness

VGG Block Structure

  • Block 1: 64 filters
    2 × Conv3×3-64 → MaxPool
  • Block 2: 128 filters
    2 × Conv3×3-128 → MaxPool
  • Block 3: 256 filters
    3 × Conv3×3-256 → MaxPool
  • Block 4: 512 filters
    3 × Conv3×3-512 → MaxPool
  • Block 5: 512 filters
    3 × Conv3×3-512 → MaxPool
  • Fully Connected
    FC-4096 → FC-4096 → FC-1000

Why 3×3 Filters Are Brilliant

7×7
One 7×7 Conv
49 parameters
1 non-linearity
5×5
Two 5×5 Conv
50 parameters
2 non-linearities
3×3
Three 3×3 Conv
27 parameters
3 non-linearities
VGG Variants
VGG-11
8 conv + 3 FC
133M params
VGG-13
10 conv + 3 FC
133M params
VGG-16
13 conv + 3 FC
138M params
VGG-19
16 conv + 3 FC
144M params
Prepared by Dr. Gorkem Kar