Explore orthogonal statistical learning in this 45-minute lecture by Vasilis Syrgkanis from Microsoft Research, presented at the Simons Institute. Delve into non-asymptotic excess risk guarantees for statistical learning with unknown nuisance parameters. Examine a two-stage sample splitting meta-algorithm and its performance under Neyman orthogonality conditions. Discover how this approach enables the use of existing statistical learning and machine learning results for new guarantees. Learn about oracle rates achievement, specific estimation algorithms, and applications in heterogeneous treatment effect estimation, offline policy optimization, domain adaptation, and learning with missing data. Investigate topics such as automated outlier removal, generalized method of moments, IV regression, neural network and forest heuristic improvements, and multi-tasking applications in 401k and gasoline demand analysis.
Overview
Syllabus
Intro
The Ease of Use of Machine Learning
The Ease of (mis)Use of Machine Learning
Making CausalML Accessible to Every Decision-Maker
Automated Outlier Removal
Desiderata for a good robust estimator
Estimation via Moment Conditions
Generalized Method of Moments
Our Results
Two-part theorem
Algorithm and Key Ingredients
Example Application Theorem: IV Regression
Experiments
Application: NLSYM
Method of Moments with Nuisance Functions
De-biased moment
The Adversarial Approach
The Direct Loss Approach
Formal Theorem
Neural Network and Forest Heuristic Improvements
Multi-Tasking
Application: 401k
Application: Gasoline Demand
Beyond Linear Moments
Taught by
Simons Institute
