Group equivariant convolutional neural networks (G-CNNs) has been introduced by Cohen & Welling (2016). These new models can be considered as an evolution of convolutional neural networks (CNNs). The key operation which G-CNNs use is G-Convolution; this is a new layer that make neural networks equivariant to new symmetries. In this paper, I’ll provide a theoretical overview of G-CNNs, G-Convolution and finally I’ll show the performance of G-CNNs vs CNNs on Fashion-MNIST and CIFAR-10 datasets.
Read the paper main.pdf for more details. In the Notebook you can have access to the code. All dependencies are intstalled in the notebook. Therefore the notebook is ready to run.
For more details about the implementation, you can look the original repository GrouPy
References:
[1] Cohen, Taco, and Max Welling. "Group equivariant convolutional networks." International conference on machine learning. PMLR, 2016.