We trained a large, deep convolutional neural network to categorize the 1.2 million high-resolution images 
from the ImageNet LSVRC-2010 competition into 1000 distinct classes. On the test set, we achieved top-1 
and top-5 error rates of 37.5% and 17.0%, significantly surpassing the previous state-of-the-art results. The 
network, with 60 million parameters and 650,000 neurons, is made up of five convolutional layers, some 
followed by max-pooling layers, and three fully connected layers with a final 1000-way softmax output. To 
accelerate training, we utilized non-saturating neurons and a highly optimized GPU implementation for the 
convolution operation. To prevent overfitting in the fully connected layers, we applied a newly developed 
regularization technique called “dropout,” which proved highly effective. We also submitted a variant of this 
model in the ILSVRC-2012 competition, achieving a top-5 test error rate of 15.3%, outperforming the 
second-place entry, which had a rate of 26.2%.