Course contentsShow
← Machine Learning and Deep Learning
Lesson 887 of 3,53820. Convolutional Neural Networks: FundamentalsFree lesson

Receptive Fields in Modern Architectures

Compare receptive field strategies in VGG (small kernels, depth) vs Inception (multi-scale) approaches.

Receptive Fields in Modern Architectures

What you'll learn: How different CNN architectures achieve large receptive fields through contrasting design philosophies—VGG's deep stacking versus Inception's multi-scale approach.

The Architectural Trade-off

Different famous CNN architectures solve the "how do we see more of the image" problem in fundamentally different ways. Remember, a larger receptive field means neurons can integrate information from broader regions of the input image—critical for recognizing complex patterns.

VGG: Go Deep with Small Kernels

VGG's strategy: Stack many 3×3 convolutions in sequence.

  • Uses exclusively small 3×3 kernels throughout the network
  • Achieves large receptive fields through depth—many sequential layers
  • Two 3×3 convs give a 5×5 receptive field; three give 7×7
  • More layers = more non-linearity = more expressive power
  • Trade-off: requires many sequential operations (slow) but simple design

Think of it like zooming out by taking many small steps backward—each step adds a little more context.

Inception: Multi-Scale Parallel Processing

Inception's strategy: Process the same input at multiple scales simultaneously.

  • Uses 1×1, 3×3, and 5×5 convolutions in parallel within the same layer
  • Also includes pooling paths
  • Concatenates all outputs, letting the network choose which scale matters
  • Achieves diverse receptive fields within a single layer
  • Trade-off: more complex architecture but captures multi-scale features efficiently

Think of it like having three photographers at different distances simultaneously—you get close-ups, medium shots, and wide angles all at once.

Why This Matters

VGG teaches us that depth with simple components works. Inception shows us that parallel multi-scale processing can be more efficient. Modern architectures often blend both philosophies—going deep and using multi-scale features.

Key Takeaway: VGG achieves large receptive fields through sequential depth with uniform small kernels, while Inception captures multiple receptive field sizes simultaneously through parallel paths—two valid solutions to the same problem.