CKConv: Continuous Kernel Convolution For Sequential Data

Open Access
Authors
Publication date 04-02-2021
Edition v1
Number of pages 17
Publisher ArXiv
Organisations
  • Faculty of Science (FNWI) - Informatics Institute (IVI)
Abstract
Conventional neural architectures for sequential data present important limitations. Recurrent networks suffer from exploding and vanishing gradients, small effective memory horizons, and must be trained sequentially. Convolutional networks are unable to handle sequences of unknown size and their memory horizon must be defined a priori. In this work, we show that all these problems can be solved by formulating convolutional kernels in CNNs as continuous functions. The resulting Continuous Kernel Convolution (CKConv) allows us to model arbitrarily long sequences in a parallel manner, within a single operation, and without relying on any form of recurrence. We show that Continuous Kernel Convolutional Networks (CKCNNs) obtain state-of-the-art results in multiple datasets, e.g., permuted MNIST, and, thanks to their continuous nature, are able to handle non-uniformly sampled datasets and irregularly-sampled data natively. CKCNNs match or perform better than neural ODEs designed for these purposes in a much faster and simpler manner.
Document type Preprint
Note Versions v2 (2021) and v3 (2022) also available on ArXiv.
Language English
Published at https://doi.org/10.48550/arXiv.2102.02611
Downloads
CKConv (Submitted manuscript)
Permalink to this page
Back