Padding Creates A "blind Spot"! Examine Those Blind Spots!

3 main points
✔️ Demonstrate how padding generates CNN artifacts (blind spots) 
✔️ Identify that uneven application of 0-padding is a solvable source of bias 
✔️ Padding is linked to CNN's foveation behavior

Mind the Pad -- CNNs can Develop Blind Spots
written by Bilal Alsallakh, Narine Kokhlikyan, Vivek Miglani, Jun Yuan, Orion Reblitz-Richardson
(Submitted on 5 Oct 2020)
Comments: Accepted at ICLR2021
Subjects: Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Machine Learning (stat.ML)
Paper  Official Code COMM Code

Introduction

Convolutional neural networks (CNNs) have become state-of-the-art feature extractors for a variety of machine learning tasks. Much work has focused on understanding the feature maps that CNN computes on input, but little attention has been paid to the spatial distribution within the feature maps. The reason for the authors' interest in this point was triggered by a mysterious case of signal detector failure. What it means is that within one frame of an on-board camera sequence, a small but visible traffic light can be detected. But in the next frame, it fails to detect the same traffic light.

It's funny. We should have been able to detect the traffic lights with high accuracy, but after only one frame, we can't. The only difference is that in the input image, the car is moving forward and so the traffic lights are shifted slightly in the vertical direction. The only difference is that in the input image, the only difference is that the signal is slightly shifted vertically because the car is moving forward. You hear a lot about CNN being invariant. Nonetheless, this research has begun because it is disconcerting that there are cases where only a slight shift makes it impossible to detect.

In conclusion, the authors attributed this to the presence of an artifact that was not a consistent feature when analyzing the spatial distribution of feature map activations (incidentally, many imaging researchers have probably seen this artifact before). In this study, we examine the causes and effects of these artifacts.

The contributions of this study are as follows

• Demonstrate how padding generates CNN artifacts
• Examine the impact of these artifacts on the task
• Identify the uneven application of 0-padding as a resolvable source of bias
• Padding is linked to CNN's foveation behavior

加藤

AI-SCHOLAR is a commentary media that introduces the latest articles on AI (artificial intelligence) in an easy-to-understand manner. The role of AI is not limited to technological innovation, as Japan's scientific capabilities are declining and the government continues to cut back on research budgets. Communicating with the world the technology of AI, its applications, and the context of the basic science that supports it is an important outreach, and can greatly influence society's understanding and impression of science. AI-SCHOLAR is designed to help eliminate the gaps in understanding of AI between the general public and experts, and to contribute to the integration of AI into society. In addition, we would like to help you embody your learning and research experiences in the media and express them in society. Anyone can explain advanced and difficult matters in difficult terms, but AI-SCHOLAR pursues "readability” and "comprehensibility" by making full use of vocabulary and design in conveying information as a medium.