Group Equivariant

Group equivariant convolutional neural networks (G-CNNs) aim to build neural networks whose outputs transform predictably under specific group actions (e.g., rotations, reflections) applied to the input. Current research focuses on improving the robustness and adaptability of G-CNNs, including developing architectures that handle partial symmetries, efficiently implement 3D transformations, and achieve invariance through techniques like triple-correlation or orthogonal moments. This field is significant because G-CNNs offer improved data efficiency, generalization, and robustness to adversarial attacks compared to traditional CNNs, with applications spanning image classification, medical image analysis, and visual correspondence.