CS5720 - Week 5
Slide 83 of 100

AlexNet: The Deep Learning Revolution

The ImageNet Moment

🏆 2012 ImageNet Competition
AlexNet achieved 15.3% top-5 error rate, crushing the second place (26.2%). This 10.8% improvement was unprecedented and marked the beginning of the deep learning era.
Why It Mattered:
• First CNN to win ImageNet
• Proved deep learning's superiority
• Sparked massive AI investment
• Changed computer vision forever
• Made GPUs essential for AI
The Team:
Alex Krizhevsky, Ilya Sutskever, and Geoffrey Hinton from the University of Toronto

Revolutionary Innovations

  • ReLU Activation
    6x faster training than tanh/sigmoid
  • 🎲
    Dropout Regularization
    Reduced overfitting dramatically
  • 🖥️
    GPU Training
    Split model across 2 GTX 580 GPUs
  • 🔄
    Data Augmentation
    Increased dataset size by 2048x
  • 📊
    Local Response Normalization
    Improved generalization (later deprecated)

LeNet-5 vs AlexNet: The Leap Forward

LeNet-5 (1998)
Input Size: 32×32×1
Layers: 7
Parameters: 60K
Activation: Tanh
Dataset: MNIST (60K)
Hardware: CPU
AlexNet (2012)
Input Size: 224×224×3
Layers: 8
Parameters: 60M (1000x more!)
Activation: ReLU
Dataset: ImageNet (1.2M)
Hardware: 2× GTX 580 GPUs
🚀 AlexNet didn't just improve on LeNet - it proved that deep learning could solve real-world problems at scale!
Prepared by Dr. Gorkem Kar