CS5720 - Week 5
Slide 82 of 100

LeNet-5: The Pioneer

Historical Context

In 1998, Yann LeCun and his team at AT&T Bell Labs introduced LeNet-5, a convolutional neural network that would lay the foundation for modern deep learning in computer vision.
The Problem: Recognizing handwritten digits for check processing

The Solution: A 7-layer convolutional neural network that could learn features automatically

The Impact: Processed millions of checks per day in real banks
Key Innovations:
• First successful CNN for real-world application
• Automatic feature learning (no hand-crafted features!)
• End-to-end trainable architecture
• Spatial hierarchy of features

Architecture Components

  • Input Layer
    32×32 grayscale images (padded MNIST)
  • C1: Convolutional Layer
    6 feature maps, 5×5 kernels → 28×28×6
  • S2: Pooling Layer
    2×2 average pooling → 14×14×6
  • C3: Convolutional Layer
    16 feature maps, 5×5 kernels → 10×10×16
  • S4: Pooling Layer
    2×2 average pooling → 5×5×16
  • C5: Convolutional Layer
    120 feature maps, 5×5 kernels → 1×1×120
  • F6 & Output
    84 units → 10 output classes

LeNet-5 Architecture Flow

Input
32×32×1
1,024 pixels
Conv
28×28×6
156 params
Pool
14×14×6
12 params
Conv
10×10×16
2,416 params
Pool
5×5×16
32 params
Conv
1×1×120
48,120 params
FC
84
10,164 params
Output
10
850 params
🏆 LeNet-5's Lasting Contributions
  • Established the Conv → Pool → FC architecture pattern
  • Proved CNNs could learn hierarchical features automatically
  • Demonstrated real-world applicability at scale
  • Inspired decades of CNN research and development
  • Total parameters: ~60,000 (tiny by today's standards!)
Prepared by Dr. Gorkem Kar