SSC Layer - A replacement for convolutional layers

Joel Rixen, Matthias Renz

Primary Area: general machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: convolutional layer, lightweight, sequence modelling

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

Convolutional layers have been used in practically every application of machine learning. We propose the SSC layer, which functions similarly to the convolutional layer but is faster, more memory efficient and competitive in terms of accuracy. The SSC layer splits the input tensor across the channel dimension, shifts each split by a different amount and subtracts the result from the input. This process enables a kernel size equal to the channel size without increasing model size, memory usage and without affecting speed, unlike convolutional layers.
The SCC layer functions in multiple dimensions and is able to replace the convolutional layer in a number of applications including image classification, sequence modelling and single-channel speech separation.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 5922