Stories
Curated narratives that connect the dots between papers, courses, and key concepts.
Interpolation Can Be Harmless
Overparameterized models can fit noisy data exactly and still predict well, yet the same extra features bleed away the signal they were meant to learn. Treating overparameterization as undersampling splits generalization into survival and contamination, and makes both calculable, for regression, classification, and beyond.
Spectrum Jails: Engineering Trust After the Fact
Cognitive radios promised flexibility that regulators could not certify line by line. Drawing on law and economics, this work asked whether identity, liability, and future sanctions could replace trust in a radio's internal design, then followed the technical tradeoffs that appear when deterrence, rights, commitment, and reputation must operate through real observations.
Witsenhausen's Counterexample: When Actions Speak
A tiny decentralized-control problem broke the linear logic of classical LQG. Connecting it to communication and information theory turned that warning into a program of constructions, converses, approximation guarantees, and new ways to see information hidden in physical action.