M12.8 CONNECT THE MECHANISM
Test what the visual model is actually using
Pet search scores 97%, then finds a "dog" on an empty lawn. Learn the four classic ways vision models fail, and flip a classifier's answer by changing each pixel one brightness level.
LESSON OVERVIEW14 min lesson
Lesson overview
Pet search scores 97%, then finds a "dog" on an empty lawn. Learn the four classic ways vision models fail, and flip a classifier's answer by changing each pixel one brightness level.
What you’ll explore
- Occlusion, background shortcuts, adversarial perturbations, and distribution shift challenge different visual assumptions; evaluation should identify conditions and consequences, not just an average score.
GO TO THE SOURCE
Original explanations, connected to the research.
Explaining and Harnessing Adversarial Examples (Goodfellow, Shlens & Szegedy, 2014)Intriguing properties of neural networks (Szegedy et al., 2013)Shortcut Learning in Deep Neural Networks (Geirhos et al., 2020)Recognition in Terra Incognita (Beery, Van Horn & Perona, 2018)Visualizing and Understanding Convolutional Networks (Zeiler & Fergus, 2013)Benchmarking Neural Network Robustness to Common Corruptions and Perturbations (Hendrycks & Dietterich, 2019)Suggest a correction
A precise note can make an explanation better.
Choose the scene and describe what needs attention. Download a feedback file to share through a channel you already use. This page does not send feedback or connect you with a reviewer.