Continuing our CTML Seminar Series is CTML GSR, Yi Li. His talk, "Targeted Deep Architectures: A TMLE-Based Framework for Robust Causal Inference in Neural Networks" will take place on October 29th at 12:00PM at Berkeley Way West, 5th Floor, Room 5401. You won't want to miss it!
Modern neural networks excel at prediction but often produce biased estimates and unreliable uncertainty for causal target parameters(e.g., average treatment effects or entire survival curves). This talk introduces Targeted Deep Architectures (TDA), a framework that embeds a targeted maximum likelihood–style update directly into a network’s parameter space. TDA freezes most weights, identifies a small “targeting” subset, and projects influence functions onto network gradients to obtain a targeting direction that iteratively removes first‑order bias. The resulting universal targeting gradient enables simultaneous debiasing of multidimensional parameters—for example, an entire survival curve—without cumbersome post‑hoc fluctuations or specialized losses.
If time permits, I will also present our new work on weighted‑path updates for simultaneous targeting of multidimensional parameters, where per‑component targeting directions are combined via statistically informed weights to produce stable, coherent updates and practical convergence criteria. The framework is model‑agnostic and integrates seamlessly with modern deep architectures.