Theory of Intelligence
DEEP LEARNING•Computer Vision•Cnn Architectures•2.2 VGG & GoogLeNet

Introduction

Two 2014 architectures with opposite philosophies: VGG's simple deep stacks of 3×3 filters, and GoogLeNet's Inception modules, which mix filter sizes and use 1×1 convolutions to stay efficient.

timer1 min read•format_list_bulletedConcept 1 of 4•forumLesson Discussion

Introduction

After AlexNet, the obvious question was: how do we go deeper? In 2014, two teams gave very different answers.

  • VGG (Oxford's Visual Geometry Group) said: keep it simple. Use only tiny 3 × 3 filters and stack lots of them.
  • GoogLeNet (Google) said: be clever. Build a special module that looks at the image with several filter sizes at once, and use 1 × 1 convolutions to keep the cost down.

Think of two ways to build a tall tower. VGG stacks identical, simple bricks very high. GoogLeNet designs a clever, lightweight prefabricated unit and stacks those. Both towers ended up much taller (deeper) than AlexNet, and both ideas are still used today.

Discuss theoretical concepts, mathematical formulas, and architectural foundations.

Supports inline code with `backticks`
No comments yet. Be the first to start the discussion!
arrow_backPrevious Lesson
2.1 LeNet & AlexNet
Next Lessonarrow_forward
2.3 ResNet & Skip Connections