Completed
Consequence: Training Modifications
Class Central Classrooms beta
YouTube videos curated by Class Central.
Classroom Contents
Why Do Our Models Learn?
Automatically move to the next video in the Classroom when playback concludes
- 1 Intro
- 2 Machine Learning Can Be Unreliable
- 3 Indeed: Machine Learning is Brittle
- 4 Backdoor Attacks
- 5 Key problem: Our models are merely (excellent!) correlation extractors Cats
- 6 Indeed: Correlations can be weird
- 7 Simple Setting: Background bias
- 8 Do Backgrounds Contain Signal?
- 9 ImageNet-9: A Fine-Grained Study Xiao Engstrom Ilyas M 2020
- 10 Adversarial Backgrounds
- 11 Background-Robust Models?
- 12 How Are Datasets Created?
- 13 Dataset Creation in Practice
- 14 Consequence: Benchmark-Task Misalignment
- 15 Prerequisite: Detailed Annotations
- 16 Ineffective Data Filtering
- 17 Multiple objects
- 18 Human-Label Disagreement
- 19 Human-Based Evaluation
- 20 Human vs ML Model Priors
- 21 Consequence: Adversarial Examples Illyas Santurkar Tsipras Engstrom Tran M 2019 (Standard) models tend to lean on "non-robust" features + Adversarial perturbations manipulate these features
- 22 Consequence: Interpretability
- 23 Consequence: Training Modifications
- 24 Robustness + Perception Alignment
- 25 Robustness + Better Representations
- 26 Counterfactual Analysis with Robust Models
- 27 ML Research Pipeline
