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.