Network-in-Network architecture

E472886

convolutional neural network architecture deep learning model architecture image classification architecture

Network-in-Network architecture is a convolutional neural network design that replaces traditional linear convolution layers with micro multilayer perceptrons (MLPs) to enhance feature abstraction and model expressiveness.

Try in SPARQL Jump to: Statements Referenced by

Statements (45)

Predicate	Object
instanceOf	convolutional neural network architecture ⓘ deep learning model architecture ⓘ image classification architecture ⓘ
aimsTo	enhance feature abstraction ⓘ improve classification performance ⓘ increase model expressiveness ⓘ
basedOn	convolutional neural networks ⓘ
characterizedBy	end-to-end training with backpropagation ⓘ nonlinear feature mapping within receptive fields ⓘ parameter efficiency compared to large fully connected layers ⓘ
citationTitle	Network In Network NERFINISHED ⓘ
domain	computer vision ⓘ deep learning research ⓘ
evaluatedOn	CIFAR-10 NERFINISHED ⓘ CIFAR-100 NERFINISHED ⓘ ImageNet (ILSVRC-2012 subset) NERFINISHED ⓘ SVHN NERFINISHED ⓘ
hasKeyIdea	perform classification with global average pooling instead of fully connected layers ⓘ replace linear filters with small neural networks ⓘ
improvesUpon	AlexNet-style CNNs NERFINISHED ⓘ
includesLayerType	convolutional layers ⓘ global average pooling layers ⓘ mlpconv layers ⓘ pooling layers ⓘ
influenced	Inception architecture NERFINISHED ⓘ design of fully convolutional networks ⓘ use of 1x1 convolutions in later CNNs ⓘ
introduces	global average pooling as a replacement for fully connected layers ⓘ mlpconv layers ⓘ
optimizationMethod	stochastic gradient descent ⓘ
outputType	class probabilities ⓘ
proposedBy	Min Lin NERFINISHED ⓘ Qiang Chen NERFINISHED ⓘ Shuicheng Yan NERFINISHED ⓘ
publicationYear	2013 ⓘ
publishedIn	arXiv:1312.4400 ⓘ
reduces	overfitting compared to large fully connected layers ⓘ
replaces	linear convolution layers with micro multilayer perceptrons ⓘ
trainingDataType	labeled images ⓘ
uses	1x1 convolutions ⓘ global average pooling ⓘ micro multilayer perceptrons ⓘ
usesActivationFunction	ReLU NERFINISHED ⓘ
usesRegularization	dropout ⓘ weight decay ⓘ

Referenced by (1)

Full triples — surface form annotated when it differs from this entity's canonical label.

Inception architecture → inspiredBy → Network-in-Network architecture ⓘ