Course contentsShow
← Machine Learning and Deep Learning
Lesson 905 of 3,53821. CNN ArchitecturesFree lesson

ResNet-34 and ResNet-50 Architecture

Comparing the plain residual block (ResNet-34) with the bottleneck design (ResNet-50).

ResNet-34 and ResNet-50 Architecture

What you'll learn: You'll understand the difference between ResNet-34's plain residual blocks and ResNet-50's bottleneck design, and why this matters for building deeper networks.

Two Flavors of ResNet

ResNet comes in several depths, but ResNet-34 and ResNet-50 represent two distinct architectural approaches that balance computational efficiency with network capacity.

ResNet-34: The Plain Residual Block

ResNet-34 uses the basic residual block we've seen before:

  • Two 3×3 convolutional layers stacked together
  • Each block maintains the same number of channels throughout
  • Example: 64 channels → 64 channels → 64 channels

This design is straightforward but becomes computationally expensive as you go deeper and increase channel counts.

ResNet-50: The Bottleneck Design

ResNet-50 introduces a bottleneck block that's more parameter-efficient:

  • Starts with a 1×1 convolution that reduces channels (e.g., 256 → 64)
  • Middle 3×3 convolution operates on the reduced dimension (64 channels)
  • Ends with a 1×1 convolution that expands back (64 → 256)

Think of it like compressing a file before sending it through a slow network connection, then decompressing it afterward. The 1×1 convolutions create a "bottleneck" that reduces computation in the expensive 3×3 layer.

Why the Switch?

The bottleneck design allows ResNet-50 to have more layers (50 vs 34) with similar or fewer parameters than ResNet-34. This makes deeper networks practical without exploding computational costs.

Both architectures use the same overall structure (initial conv layer, four stages with increasing channels, global average pooling), but differ in their building blocks.

Key Takeaway: ResNet-34 uses simple two-layer blocks, while ResNet-50 uses three-layer bottleneck blocks that compress and expand channels for efficiency—enabling deeper networks without proportional cost increases.