Course contentsShow
← Machine Learning and Deep Learning
Lesson 863 of 3,53820. Convolutional Neural Networks: FundamentalsFree lesson

Common Filter Sizes: 3x3, 5x5, 1x1

Standard kernel sizes used in practice and their trade-offs.

Common Filter Sizes: 3x3, 5x5, 1x1

What you'll learn: Why certain kernel sizes dominate CNN architectures and how to choose between them.

Why These Sizes?

While you can technically use any filter size, three dimensions have become standard in practice: 3×3, 5×5, and 1×1. Each offers different trade-offs between receptive field, computational cost, and feature complexity.

The 3×3 Filter: The Workhorse

The 3×3 filter is the most common choice in modern CNNs. It's small enough to be computationally efficient yet large enough to capture meaningful local patterns. Think of it as looking at a small neighborhood around each pixel—just enough context without being wasteful.

Key advantage: Two stacked 3×3 convolutions give you the same receptive field as one 5×5 filter but with fewer parameters (18 vs 25 per channel) and more non-linearity.

The 5×5 Filter: Broader Context

5×5 filters capture larger spatial patterns in a single operation. They're useful early in networks when input resolution is high and you want to quickly capture broader features.

Trade-off: They have 2.78× more parameters than 3×3 (25 vs 9), making them computationally heavier. This is why modern architectures often replace them with stacked 3×3 layers.

The 1×1 Filter: The Channel Mixer

The 1×1 filter might seem odd—it doesn't look at neighboring pixels at all! Instead, it operates across channels at each spatial location, mixing channel information. Think of it as a per-pixel fully connected layer.

Uses: Reducing/expanding channel dimensions, adding non-linearity without spatial mixing, and creating "bottleneck" layers that reduce parameters.

Key Takeaway: 3×3 filters dominate modern CNNs for efficiency, 5×5 captures broader patterns at higher cost, and 1×1 filters manipulate channel dimensions without spatial convolution—each size serves distinct architectural purposes.