Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification
Level 5 - mechanism / opinion, no new human data
Level 5 by design analogy (computational algorithm development and benchmark evaluation without human clinical data).
OpenAlex W2949608135 · doi:10.48550/arxiv.1502.01852
What was done
The authors investigated rectified neural networks for image classification by proposing a Parametric Rectified Linear Unit (PReLU) to generalize traditional activation units and deriving a robust initialization method tailored for rectifier nonlinearities to enable training deep networks from scratch. The method was evaluated on the ImageNet 2012 classification dataset.
What was found
Networks utilizing PReLU and the new initialization achieved a 4.94% top-5 test error on the ImageNet 2012 classification dataset, reflecting a 26% relative improvement over the ILSVRC 2014 winner (GoogLeNet at 6.66%) and outperforming reported human-level performance of 5.1%.
Why it matters
This work introduced robust weight initialization and learnable activation functions that enabled the reliable training of very deep convolutional networks from scratch, achieving a milestone in computer vision benchmark performance.
Limits
The findings are derived from a single standardized computational benchmark (ImageNet 2012) without testing on other visual domains in the abstract. Exact sample sizes and performance on downstream real-world vision tasks were not reported.
Cited by
- supports Human benchmark error rate on the ImageNet 1,000-class object recognition task was measured at roughly 4%.