AlexNet

E74105

convolutional neural network architecture deep learning model image classification model

AlexNet is a pioneering deep convolutional neural network architecture that dramatically advanced image recognition performance and helped spark the modern deep learning revolution after winning the 2012 ImageNet competition.

Aliases (1)

ImageNet classification with deep convolutional neural networks ×1

Statements (49)

Predicate	Object
instanceOf	convolutional neural network architecture → deep learning model → image classification model →
achievement	won the ILSVRC 2012 image classification task →
baselineImprovement	significantly reduced top-5 error compared to previous state of the art →
category	supervised learning model →
competition	ImageNet Large Scale Visual Recognition Challenge →
competitionOrganizer	ImageNet →
competitionYear	2012 →
convolutionalLayerCount	5 →
countryOfDevelopment	Canada →
dataset	ImageNet →
developer	Alex Krizhevsky → Geoffrey Hinton → Ilya Sutskever →
field	computer vision → deep learning →
framework	CUDA-based custom implementation →
fullyConnectedLayerCount	3 →
influenced	GoogLeNet → ResNet → VGGNet → modern deep convolutional network design →
inputColorChannels	3 →
inputImageSize	224x224 pixels (approximately) →
institution	University of Toronto →
introducedInPaper	ImageNet Classification with Deep Convolutional Neural Networks →
layerCount	8 learned layers →
lossFunction	cross-entropy loss →
notableContribution	demonstrated effectiveness of deep CNNs on large-scale image recognition → popularized use of GPUs for deep learning → showed benefits of ReLU over tanh and sigmoid in deep networks →
optimization	backpropagation →
outputClasses	1000 →
parameterCount	approximately 60 million parameters →
publicationYear	2012 →
significance	sparked the modern deep learning revolution in computer vision →
top5ErrorRate	15.3% →
trainingDatasetSize	over 1 million images →
trainingHardware	GPU → NVIDIA GTX 580 →
trainingTechnique	stochastic gradient descent with momentum →
trainingTime	several days →
usesActivationFunction	ReLU →
usesNonlinearity	rectified linear units →
usesNormalization	local response normalization →
usesPooling	max pooling →
usesRegularization	data augmentation → dropout →

Referenced by (2)

Subject (surface form when different)	Predicate
LeNet →	influenced
Ilya Sutskever ("ImageNet classification with deep convolutional neural networks") →	notableWork