Learn Before
Deeper Bottleneck Architectures - Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor
ImageNet Classification and Model Variations - Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor
Deep ResNet Scaling with Bottleneck Blocks
Bottleneck blocks enable the construction of substantially deeper networks, specifically 50-layer, 101-layer, and 152-layer ResNets:
- ResNet-50: Created by replacing each 2-layer block of the 34-layer network with the 3-layer bottleneck block, utilizing projection shortcuts (Option B) for increasing dimensions and requiring FLOPs.
- ResNet-101 and ResNet-152: Built by stacking significantly more 3-layer blocks, particularly across feature maps of size and .
Even with 152 layers, ResNet-152 requires FLOPs, maintaining lower computational complexity than VGG-16 ( FLOPs) and VGG-19 ( FLOPs). Across these configurations, the networks show no degradation problem and yield substantial accuracy improvements as depth increases.
0
1
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
Deeper Bottleneck Architectures - Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor
Ch.4 Residual Network Experiments and Applications - Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor
ImageNet Classification and Model Variations - Foundational Deep Learning Architectures: Transformers and Residual Networks @ University of Michigan - Ann Arbor
Learn After
Match each network architecture to its corresponding computational complexity (FLOPs) or structural design specification.
How should the team address these two concerns using the known computational complexity and scaling properties of bottleneck ResNets?
How is the ResNet-50 architecture structurally derived from the 34-layer ResNet baseline?
Scaling network depth across 50-layer, 101-layer, and 152-layer ResNets results in the degradation problem and lower accuracy.
Across which specific feature map sizes are significantly more 3-layer blocks stacked to build ResNet-101 and ResNet-152?
Compare the computational complexity of ResNet-152 with VGG-16 and VGG-19, and describe the impact of scaling to this depth on model degradation and accuracy.