Kaiming He · 2015 · computational benchmark and algorithm development study · n=?

Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification

Cited 18994 times in the scientific literature.

Level 5 - mechanism / opinion, no new human data

Level 5 by design analogy; non-clinical computer science and machine learning benchmark experiment.

OpenAlex W1677182931 · doi:10.1109/iccv.2015.123 · record verified 2026-08-26

What was done

The authors investigated rectified neural networks for image classification by introducing two techniques: Parametric Rectified Linear Units (PReLU), a learnable generalization of traditional rectified units with negligible extra computational cost, and a robust initialization method derived specifically for rectifier non-linearities to facilitate training extremely deep architectures from scratch. The methods were evaluated on the ImageNet 2012 classification dataset.

What was found

Using PReLU and the derived initialization method, the model achieved a 4.94% top-5 error rate on the ImageNet 2012 test set. This was a 26% relative improvement over the ILSVRC 2014 winner (GoogLeNet, 6.66%) and surpassed the cited human-level performance baseline of 5.1% on this benchmark.

Why it matters

This work introduced PReLU and robust weight initialization (He initialization) for deep networks, demonstrating that deep rectified networks can surpass human-level performance on large-scale visual classification benchmarks.

Limits

Findings are limited to computer vision benchmarking on the ImageNet 2012 dataset; performance on other tasks, architectures, or modalities was not reported in the abstract. Comparisons to human performance depend on a single reported baseline rather than direct experimental testing against humans across varied conditions.

Cited by