Concept icon
Concept

Bottleneck Residual Blocks

To build extremely deep networks such as ResNet-50, ResNet-101, and ResNet-152 within feasible training times, the standard two-layer residual block is modified into a three-layer bottleneck block.

In a bottleneck block, the residual function F\mathcal{F} consists of a stack of three convolutions: 1×11 \times 1, 3×33 \times 3, and 1×11 \times 1. The first 1×11 \times 1 convolution is responsible for reducing the channel dimensions (e.g., from 256 to 64), the middle 3×33 \times 3 convolution operates on this lower-dimensional bottleneck, and the final 1×11 \times 1 convolution restores the dimensions back to the high-dimensional representation (e.g., restoring back to 256).

Parameter-free identity shortcuts are essential for bottleneck architectures. Because the shortcut connects the two high-dimensional ends of the block, replacing the identity shortcut with a projection shortcut would double both the parameter count and time complexity. By maintaining parameter-free identity shortcuts, deep networks like the 152-layer ResNet achieve higher accuracy than shallower baselines while retaining lower computational complexity (11.3 billion FLOPs) than VGG-16 and VGG-19.

0

1

Concept icon
Updated 2026-09-07

Tags

Prep Sessions

Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor

Ch.3 Deep Residual Network Architecture - Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor

Residual Formulation and Shortcut Connections - Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor

Identity versus Projection Shortcuts - Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor

Deeper Bottleneck Architectures - Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor