Deep Rotation Equivariant Network
This work addresses a computational bottleneck in rotation-equivariant deep learning, offering a more efficient solution for applications requiring robustness to rotations, though it is incremental in nature.
The paper tackles the computational inefficiency of existing rotation-equivariant networks by proposing Deep Rotation Equivariant Network (DREN) with new layers that apply transformations to filters instead of feature maps, achieving over 2x speedup and reduced memory overhead while improving performance on Rotated MNIST and CIFAR-10 datasets.
Recently, learning equivariant representations has attracted considerable research attention. Dieleman et al. introduce four operations which can be inserted into convolutional neural network to learn deep representations equivariant to rotation. However, feature maps should be copied and rotated four times in each layer in their approach, which causes much running time and memory overhead. In order to address this problem, we propose Deep Rotation Equivariant Network consisting of cycle layers, isotonic layers and decycle layers. Our proposed layers apply rotation transformation on filters rather than feature maps, achieving a speed up of more than 2 times with even less memory overhead. We evaluate DRENs on Rotated MNIST and CIFAR-10 datasets and demonstrate that it can improve the performance of state-of-the-art architectures.