FlexConv: Continuous Kernel Convolutions with Differentiable Kernel Sizes

FlexConv: Continuous Kernel Convolutions with Differentiable Kernel Sizes

Romero, David W. (Vrije Universiteit Amsterdam)
Bruintjes, R. (TU Delft Pattern Recognition and Bioinformatics)
Bekkers, Erik J. (Universiteit van Amsterdam)
Tomczak, Jakub M. (Vrije Universiteit Amsterdam)
Hoogendoorn, Mark (Vrije Universiteit Amsterdam)
van Gemert, J.C. (TU Delft Pattern Recognition and Bioinformatics)

2022

When designing Convolutional Neural Networks (CNNs), one must select the size of the convolutional kernels before training. Recent works show CNNs benefit from different kernel sizes at different layers, but exploring all possible combinations is unfeasible in practice. A more efficient approach is to learn the kernel size during training. However, existing works that learn the kernel size have a limited bandwidth. These approaches scale kernels by dilation, and thus the detail they can describe is limited. In this work, we propose FlexConv, a novel convolutional operation with which high bandwidth convolutional kernels of learnable kernel size can be learned at a fixed parameter cost. FlexNets model long-term dependencies without the use of pooling, achieve state-of-the-art performance on several sequential datasets, outperform recent works with learned kernel sizes, and are competitive with much deeper ResNets on image benchmark datasets. Additionally, FlexNets can be deployed at higher resolutions than those seen during training. To avoid aliasing, we propose a novel kernel parameterization with which the frequency of the kernels can be analytically controlled. Our novel kernel parameterization shows higher descriptive power and faster convergence speed than existing parameterizations. This leads to important improvements in classification accuracy.

http://resolver.tudelft.nl/uuid:4ea7e75d-3cc5-4e5a-b051-168c6f80f4b5

10th International Conference on Learning Representations, 2022-04-25 → 2022-04-29, virtual event

Institutional Repository

conference paper

© 2022 David W. Romero, R. Bruintjes, Erik J. Bekkers, Jakub M. Tomczak, Mark Hoogendoorn, J.C. van Gemert