ImageNet Classification Results and Benchmarks
On the ImageNet 2012 classification benchmark (1,000 classes, 1.28 million training images, and 50k validation images), scaling depth with residual bottleneck networks produces substantial, monotonic accuracy gains without exhibiting degradation. In 10-crop evaluation on the validation set, top-1 and top-5 error rates steadily decrease as depth scales up:
- ResNet-50: 22.85% top-1 error, 6.71% top-5 error
- ResNet-101: 21.75% top-1 error, 6.05% top-5 error
- ResNet-152: 21.43% top-1 error, 5.71% top-5 error
Under single-model evaluation using dense multi-scale testing, ResNet-152 achieves a top-5 validation error of 4.49% (and 19.38% top-1 error). This standalone model outperforms previous state-of-the-art ensemble systems, including GoogLeNet (6.66% test top-5), VGG ensembles (7.32% test top-5), and BN-inception (4.82% test top-5).
Combining six models of varying depths (incorporating only two 152-layer models at the submission deadline) forms an ensemble that achieves a 3.57% top-5 error on the ImageNet evaluation test set. This performance earned first place in the ILSVRC 2015 classification competition.
0
1
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
Learn After
Match each ResNet architecture depth to its corresponding top-1 error rate achieved during 10-crop evaluation on the ImageNet validation set.
Under single-model evaluation using dense multi-scale testing, ResNet-152 achieves a top-5 validation error of ___%.
Describe the composition of the winning ILSVRC 2015 ResNet ensemble and state the top-5 error rate it attained on the evaluation test set.