In this blog, group equivariant convolutional networks, or G-CNNs, are explained. In 2016, T. S. Cohen and M. Welling published an article that both provided the concept and its evaluation which, at the time, resulted in state-of-the-art performances on some well known data sets, including CIFAR10. In doing so, they designed layers that were equivariant to unaddressed relevant isometries of the sampling lattice for which conventional CNN’s exploit translations only. My hope is to provide the reader with a more detailed explanation of their findings by the addition of an improved visual interpretation.
Whole blog? Click here.