Convolution Can Incur Foveation Effects

Rethinking ML Papers Workshop at ICLR

Abstract

We demonstrate how boundary treatment in convolutional networks can incur foveation effects: Impacted pixels have fewer ways to contribute to the computation than central pixels. Different padding mechanisms can either eliminate or aggravate these effects (Alsallakh et al. 2021). This is made obvious via a web-based interactive visualization, available at https://mind-the-pad.github.io/

Latest Publications

Boosted Dense Retriever

Patrick Lewis, Barlas Oğuz, Wenhan Xiong, Fabio Petroni, Wen-tau Yih, Sebastian Riedel

NAACL - 2022