ImageNet classification with deep convolutional neural networks
Level 5 - mechanism / opinion, no new human data
Level 5 by design analogy; non-clinical computational algorithm development and benchmark testing without human subjects.
OpenAlex W2163605009 · doi:10.1145/3065386
What was done
Authors developed and trained a deep convolutional neural network consisting of 60 million parameters, 650,000 neurons, five convolutional layers (some followed by max-pooling), and three fully connected layers ending in a 1000-way softmax. The model was trained on 1.2 million high-resolution images across 1,000 classes from the ImageNet LSVRC-2010 dataset using non-saturating neurons, GPU-accelerated convolution operations, and dropout regularization in fully connected layers. A variant was also evaluated in the ILSVRC-2012 competition.
What was found
On the ImageNet LSVRC-2010 test dataset, the model achieved top-1 and top-5 error rates of 37.5% and 17.0%, respectively. In the ILSVRC-2012 competition, a model variant achieved a winning top-5 test error rate of 15.3%, outperforming the second-place entry which achieved a 26.2% top-5 error rate.
Why it matters
This work demonstrated that large-scale deep convolutional neural networks trained on GPUs with dropout regularization substantially outperform traditional computer vision pipelines on complex visual classification tasks.
Limits
The abstract reports results exclusively on static ImageNet classification benchmarks without evaluating performance on out-of-distribution imagery, real-time inference constraints, or broader computer vision tasks like object detection and segmentation. Computational hardware requirements and energy costs for training 60 million parameters are not detailed in the abstract.
Cited by
- supports In the 2012 ImageNet Challenge, the neural network model AlexNet reduced the image classification error rate down to the teens.