Padding Aware Neurons

TL;DR

This paper introduces Padding Aware Neurons (PANs), a type of filter in convolutional models that recognize input border locations, revealing a spatial bias induced by static padding, with implications for model performance and safety.

Contribution

The paper identifies and characterizes PANs in convolutional models, proposing a method to detect them and analyzing their impact on model biases and performance.

Findings

PANs are present in all tested models.

PANs induce strong border-related biases.

Padding and PANs significantly affect model behavior.

Abstract

Convolutional layers are a fundamental component of most image-related models. These layers often implement by default a static padding policy (\eg zero padding), to control the scale of the internal representations, and to allow kernel activations centered on the border regions. In this work we identify Padding Aware Neurons (PANs), a type of filter that is found in most (if not all) convolutional models trained with static padding. PANs focus on the characterization and recognition of input border location, introducing a spatial inductive bias into the model (e.g., how close to the input's border a pattern typically is). We propose a method to identify PANs through their activations, and explore their presence in several popular pre-trained models, finding PANs on all models explored, from dozens to hundreds. We discuss and illustrate different types of PANs, their kernels and behaviour. To understand their relevance, we test their impact on model performance, and find padding and PANs to induce strong and characteristic biases in the data. Finally, we discuss whether or not PANs are desirable, as well as the potential side effects of their presence in the context of model performance, generalisation, efficiency and safety.