4/2/25 Seminar: "Improving Finite Sample Performance in Auto-Debiased Causal Neural Networks with RieszDragon"

Our CTML Seminar Series continues on April 2nd with an exciting talk led by Nolan Gunter on "Improving Finite Sample Performance in Auto-Debiased Causal Neural Networks with RieszDragon." Don't miss this talk taking place at 12:00PM at Berkeley Way West, 5th Floor, Room 5401.

The Riesz representation theorem allows us to express any target parameter as an inner product of a conditional mean and the Riesz representer, sparking new causal inference work to directly estimate the Riesz representer instead of solving for its analytical form. Causal neural networks are a popular choice for double machine learning Riesz regression estimators because they are universal function approximators. However, existing structures poorly estimate the Riesz representer and are susceptible to under-coverage or variance blow up under positivity violations. We propose RieszDragon, a neural network architecture combining Riesz-based regression weighting and collaborative, stratified potential outcome modeling to improve MSE and coverage in finite samples.