Explore the intricacies of machine learning features in this thought-provoking lecture by Aleksander Madry from MIT. Delve into the success of deep learning and examine the key phenomenon of adversarial perturbations. Analyze machine learning through the lens of adversarial robustness, comparing human and ML model perspectives. Investigate the robust features model and its implications for data efficiency, perception alignment, and image synthesis. Discover how robustness can lead to better representations and explore the potential of counterfactual analysis with robust models. Gain insights into why adversarial examples arise from non-robust features in data and consider the broader implications for the field of machine learning.
Overview
Syllabus
Intro
Machine Learning: A Success Story
Why Do We Love Deep Learning?
Key Phenomenon: Adversarial Perturbations
ML via Adversarial Robustness Lens
But: "How"/"what" does not tell us "why"
Why Are Adv. Perturbations Bad?
Human Perspective
ML Perspective
A Simple Experiment
The Robust Features Model
The Simple Experiment: A Second Look
Human vs ML Model Priors
New capability: Robustification
Some Direct Consequences
Robustness and Data Efficiency
Robustness + Perception Alignment
Robustness → Better Representations
Robustness + Image Synthesis
Problem: Correlations can be weird
Useful tool(?): Counterfactual Analysis with Robust Models
Adversarial examples arise from non-robust features in the data
Taught by
Institute for Advanced Study
