Course contentsShow
← Machine Learning and Deep Learning
Lesson 894 of 3,53821. CNN ArchitecturesFree lesson

GoogLeNet and the Inception Module

Learn how GoogLeNet used Inception modules to efficiently compute multi-scale features in parallel.

GoogLeNet and the Inception Module

You'll learn how GoogLeNet revolutionized CNN design by computing multi-scale features in parallel using Inception modules.

The Core Innovation

While VGGNet went deeper with simple stacked layers, GoogLeNet (also called Inception v1) took a different approach in 2014: what if we processed features at multiple scales simultaneously?

Instead of choosing between a 3x3 or 5x5 convolution, the Inception module does both at once—plus 1x1 convolutions and pooling—then concatenates the results. This allows the network to "look" at the same input through different-sized windows in parallel.

How an Inception Module Works

Think of it like analyzing a photograph with multiple magnifications simultaneously:

  • 1x1 convolutions: Capture fine, point-wise features
  • 3x3 convolutions: Detect small local patterns
  • 5x5 convolutions: Recognize larger structures
  • Max pooling branch: Preserve spatial information

All branches process the same input, and their outputs are concatenated along the channel dimension.

Computational Efficiency

The genius is using 1x1 convolutions as bottlenecks before expensive 3x3 and 5x5 operations. This dramatically reduces parameters compared to naive parallel convolutions—remember, 1x1 convs reduce channel dimensions cheaply (as you learned previously).

GoogLeNet achieved 22 layers deep with 12x fewer parameters than AlexNet while being more accurate!

Architecture Highlights

GoogLeNet stacks multiple Inception modules, uses global average pooling instead of fully connected layers (eliminating millions of parameters), and introduced auxiliary classifiers during training to combat vanishing gradients.

Key Takeaway: GoogLeNet's Inception module processes multi-scale features in parallel using different kernel sizes, achieving efficiency through 1x1 bottleneck convolutions before computationally expensive operations.