Explore the intricacies of machine learning model behavior in this thought-provoking lecture by MIT Professor Aleksander Madry. Delve into the alignment between benchmark-driven ML paradigms and real-world applications, examining biases in datasets like ImageNet and how state-of-the-art models exploit them. Discover how these biases stem from data collection and curation processes, and learn to quantify them using standard tools. Gain insights into the challenges of deploying reliable and responsible AI in real-world scenarios, covering topics such as background bias, adversarial examples, and the consequences of benchmark-task misalignment. Understand the implications for model interpretability, training modifications, and the overall machine learning research pipeline.
Overview
Syllabus
Intro
Machine Learning Can Be Unreliable
Indeed: Machine Learning is Brittle
Backdoor Attacks
Key problem: Our models are merely (excellent!) correlation extractors Cats
Indeed: Correlations can be weird
Simple Setting: Background bias
Do Backgrounds Contain Signal?
ImageNet-9: A Fine-Grained Study Xiao Engstrom Ilyas M 2020
Adversarial Backgrounds
Background-Robust Models?
How Are Datasets Created?
Dataset Creation in Practice
Consequence: Benchmark-Task Misalignment
Prerequisite: Detailed Annotations
Ineffective Data Filtering
Multiple objects
Human-Label Disagreement
Human-Based Evaluation
Human vs ML Model Priors
Consequence: Adversarial Examples Illyas Santurkar Tsipras Engstrom Tran M 2019 (Standard) models tend to lean on "non-robust" features + Adversarial perturbations manipulate these features
Consequence: Interpretability
Consequence: Training Modifications
Robustness + Perception Alignment
Robustness + Better Representations
Counterfactual Analysis with Robust Models
ML Research Pipeline
Taught by
MITCBMM
