An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling

The paper questions the common assumption that sequence modeling is synonymous with recurrent networks, noting recent results where convolutional architectures outperform recurrent ones on tasks such as audio synthesis and machine translation. To answer which architecture to use for a new task, the authors conduct a systematic evaluation of generic convolutional and recurrent architectures, testing them across a broad range of standard tasks and datasets commonly used to benchmark recurrent networks.

Their results show that a simple convolutional architecture outperforms canonical recurrent networks such as LSTMs across a diverse range of tasks and datasets, while also demonstrating longer effective memory. They conclude that the association between sequence modeling and recurrent networks should be reconsidered and that convolutional networks should be regarded as a natural starting point for sequence modeling tasks. Code was made available to support related work.