Comparison

Plain versus Residual Baseline Architectures

The ImageNet plain baselines are primarily inspired by VGG networks. Their convolutional layers mostly use 3×33\times3 filters, with the same number of filters at a given feature-map size and twice as many filters whenever the feature-map size is halved. Downsampling uses stride-2 convolutions. Global average pooling is followed by a 1000-way fully connected layer and softmax. The 34-layer plain baseline requires 3.6 billion FLOPs, about 18% of VGG-19's 19.6 billion FLOPs.

Increasing the plain network from 18 to 34 layers raises validation top-1 error from 27.94% to 28.54%, illustrating the degradation problem. This optimization difficulty was not attributed to vanishing gradients: Batch Normalization was used, and forward signal variances and backward gradient norms remained healthy. Residual counterparts add shortcut connections across pairs of 3×33\times3 filters, using identity shortcuts and zero-padding when dimensions change. ResNet-34 achieves 25.03% top-1 error, compared with 27.88% for ResNet-18 and 28.54% for the 34-layer plain network. Its lower training error throughout optimization demonstrates that the residual formulation allowed the tested deeper network to benefit from added depth.

0

1

Updated 2026-09-19

Tags

Prep Sessions

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