Course contentsShow
← Machine Learning and Deep Learning
Lesson 882 of 3,53820. Convolutional Neural Networks: FundamentalsFree lesson

Impact of Stride on Receptive Fields

Understand how stride increases receptive field growth rate while reducing spatial dimensions.

Impact of Stride on Receptive Fields

What You'll Learn

You'll learn how increasing stride accelerates receptive field growth while simultaneously shrinking your spatial dimensions.

How Stride Affects Receptive Field Growth

In previous lessons, you learned how to calculate receptive fields and that they grow as you stack layers. Now we'll see how stride acts as a multiplier for that growth.

When you use stride > 1, the receptive field expands more rapidly with each additional layer. Here's why: stride determines how many input pixels your sliding window skips. If stride = 2, each output position "looks at" input positions that are 2 pixels apart instead of 1. When you stack another layer on top, that 2x spacing compounds—the second layer's kernel spans positions that are already spaced out, multiplying the effective coverage.

The Trade-off

Think of stride like zooming out on a map. With stride = 2:

  • Receptive field grows faster: Each layer covers more territory in the original image
  • Spatial dimensions shrink: You get fewer output positions (half the width/height with stride 2)

This is the fundamental trade-off: you gain a larger "field of view" per neuron, but you lose spatial resolution. A 3x3 kernel with stride 2 increases receptive field contributions more aggressively than stride 1, because each step in the output corresponds to a 2-pixel jump in the input.

When This Matters

Understanding this trade-off helps you design efficient architectures. Want to quickly build a large receptive field to capture global context? Use strided convolutions. Need to preserve fine-grained spatial details? Stick with stride 1 and use pooling more selectively.

Key Takeaway: Stride greater than 1 accelerates receptive field growth proportionally while reducing spatial dimensions, creating a fundamental trade-off between coverage and resolution.