Exploiting Cyclic Symmetry in Convolutional Neural Networks
arXiv:1602.02660
Abstract
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.
10 pages, 6 figures, accepted for publication at ICML 2016
References in corpus (4)
Cited by in corpus (48)
- Informed Machine Learning -- A Taxonomy and Survey of Integrating Knowledge into Learning Systems
- Group Equivariant Convolutional Networks
- Retinal vessel segmentation based on Fully Convolutional Neural Networks
- Layer-wise training convolutional neural networks with smaller filters for human activity recognition using wearable sensors
- PointCNN: Convolution On -Transformed Points
- Equivariance Through Parameter-Sharing
- Incorporating Symmetry into Deep Dynamics Models for Improved Generalization
- PDE-based Group Equivariant Convolutional Neural Networks
- A General Theory of Equivariant CNNs on Homogeneous Spaces
- Equivariant Transformer Networks
- CyCNN: A Rotation Invariant CNN using Polar Mapping and Cylindrical Convolution Layers
- Doubly Convolutional Neural Networks
- Transformation Consistent Self-ensembling Model for Semi-supervised Medical Image Segmentation
- Group Convolutional Neural Networks Improve Quantum State Accuracy
- Provably scale-covariant continuous hierarchical networks based on scale-normalized differential expressions coupled in cascade
- Scale-invariant scale-channel networks: Deep networks that generalise to previously unseen scales
- Hexagonal Image Processing in the Context of Machine Learning: Conception of a Biologically Inspired Hexagonal Deep Learning Framework
- A Data and Compute Efficient Design for Limited-Resources Deep Learning
- B-Spline CNNs on Lie Groups
- Coordinate Independent Convolutional Networks -- Isometry and Gauge Equivariant Convolutions on Riemannian Manifolds
- Spin-Weighted Spherical CNNs
- Modularity in Deep Learning: A Survey
- Quantum algorithms for group convolution, cross-correlation, and equivariant transformations
- Self-Supervised Multi-View Learning via Auto-Encoding 3D Transformations
- Relative stability toward diffeomorphisms indicates performance in deep nets
- E(n) Equivariant Normalizing Flows
- Learning Equivariant Energy Based Models with Equivariant Stein Variational Gradient Descent
- Trajectory Prediction using Equivariant Continuous Convolution
- IPC-Net: 3D point-cloud segmentation using deep inter-point convolutional layers
- GraphTER: Unsupervised Learning of Graph Transformation Equivariant Representations via Auto-Encoding Node-wise Transformations
- Spectral Roll-off Points Variations: Exploring Useful Information in Feature Maps by Its Variations
- Learning Compressed Transforms with Low Displacement Rank
- Invariance Measures for Neural Networks
- NeRV: Neural Representations for Videos
- Automatic Analysis System of Calcaneus Radiograph: Rotation-Invariant Landmark Detection for Calcaneal Angle Measurement, Fracture Identification and Fracture Region Segmentation
- The Convolution Exponential and Generalized Sylvester Flows
- Learnable Gabor modulated complex-valued networks for orientation robustness
- Rotaflip: A New CNN Layer for Regularization and Rotational Invariance in Medical Images
- Circular-Symmetric Correlation Layer based on FFT
- Euclidean Invariant Recognition of 2D Shapes Using Histograms of Magnitudes of Local Fourier-Mellin Descriptors
- Equivariant neural networks for inverse problems
- Can we have it all? On the Trade-off between Spatial and Adversarial Robustness of Neural Networks
- Equivariant Deep Dynamical Model for Motion Prediction
- Learning Equivariant Representations
- Self-Refining Deep Symmetry Enhanced Network for Rain Removal
- Group Equivariant Subsampling
- On Universalized Adversarial and Invariant Perturbations
- Rotation Equivariant 3D Hand Mesh Generation from a Single RGB Image