Atlas record: https://www.aitimeline.dev/en/events/evt-0007/
Benchmark · 2012
A deep convolutional network reduced ImageNet error.
Its benchmark result helped make deep convolutional training a central vision approach.
The result is specific to the ImageNet challenge and its evaluation setup.
NeurIPS · Published 2012
ImageNet (organisers of the challenge)
AlexNet in 2012 uses this nonlinearity.
The competition AlexNet won in 2012.
The AlexNet paper trains and evaluates its network on ImageNet images: direct use of the earlier dataset.
The recipe AlexNet was trained by three years later.
The AlexNet paper describes training on GPUs using this platform.
The features here are hand-designed; convolutional networks later learn them.
LeNet-5 and the MNIST set are the direct basis for AlexNet.
AlexNet in 2012 is the same architecture, enlarged and moved onto a graphics processor.
Both recognise images through layers of local features; the later one ran on much larger data and on GPUs.
Two ways of proving the same thing in the same year: scale produces features.
The AlexNet paper (section 4.2) calls dropout a recently introduced technique, applies it with probability 0.5 in the first two fully connected layers and says that without it the network overfits substantially.
R-CNN takes the convolutional network of Krizhevsky and colleagues, pre-trained on ILSVRC 2012, and fine-tunes it on regions for detection.
The state is given as screen pixels through a convolutional network.
Architectures like AlexNet became available as ready models with weights.
The paper names the deep convolutional networks of Krizhevsky and colleagues as what made its approach possible.
Both concern deep convolutional networks for images; the later one removes a limit on depth.
The model is built on convolutions, causal and dilated ones.
The convolution that ruled vision from 2012 turned out to be optional given enough data.