Exploiting Cyclic Symmetry in Convolutional Neural Networks
This work addresses the need for more efficient parameter use in convolutional neural networks for image tasks with rotational symmetry, though it is incremental as it builds on existing equivariance concepts.
The paper tackled the problem of rotational symmetry in images by introducing four architectural operations that encode rotation equivariance into neural networks, enabling parameter sharing across orientations and resulting in improved performance with smaller models on three datasets.
Many classes of images exhibit rotational symmetry. Convolutional neural networks are sometimes trained using data augmentation to exploit this, but they are still required to learn the rotation equivariance properties from the data. Encoding these properties into the network architecture, as we are already used to doing for translation equivariance by using convolutional layers, could result in a more efficient use of the parameter budget by relieving the model from learning them. We introduce four operations which can be inserted into neural network models as layers, and which can be combined to make these models partially equivariant to rotations. They also enable parameter sharing across different orientations. We evaluate the effect of these architectural modifications on three datasets which exhibit rotational symmetry and demonstrate improved performance with smaller models.