CS5720 - Week 4
Slide 76 of 80

CNN Parameter Sharing

What is Parameter Sharing?

Parameter sharing in CNNs means using the same filter (weights) across all spatial locations of the input. One filter detects the same pattern everywhere in the image.
How it works:

• One 3×3 filter = 9 parameters
• Applied to entire 224×224 image
• Same 9 parameters used everywhere
• Dramatically reduces parameter count
🔍 Think of it like...
Using the same magnifying glass to examine every part of a large document. You don't need a different magnifying glass for each page!

Benefits of Parameter Sharing

  • Memory Efficiency
    Massive reduction in parameter count and memory usage
  • 🔄
    Translation Invariance
    Detects patterns regardless of position in image
  • 🎯
    Better Generalization
    Reduces overfitting by constraining parameters
  • 💨
    Faster Training
    Fewer parameters mean faster forward and backward passes

Parameter Count: With vs Without Sharing

Without Parameter Sharing
150M+
Parameters
W1
W2
W3
W4
W5
W6
W7
W8
W9
W10
W11
W12
W13
W14
W15
W16
Every spatial location has unique weights
Memory: ~600 MB
Training time: Very slow
With Parameter Sharing
25K
Parameters
W
W
W
W
W
W
W
W
W
W
W
W
W
W
W
W
Same weights shared everywhere
Memory: ~100 KB
Training time: Fast
Parameter sharing reduces memory usage by 6000× in this example!
Prepared by Dr. Gorkem Kar