"Designing Principled ML Algorithms via Modularity" – Dhruv Rohatgi, Talks at TTIC

“Designing Principled ML Algorithms via Modularity” Dhruv Rohatgi, MIT Originally recorded on March 31, 2026, at TTIC. In this talk, Dhruv Rohatgi presents a modular approach to machine learning theory, using existing components such as classifiers, regressors, and generative models as building blocks for more complex systems. He applies this framework across stages of the language model pipeline—including fine-tuning, reinforcement learning, and generation—highlighting how principled algorithm design can mitigate challenges like the “curse of horizon” in sequential decision-making. Timestamps: 00:00 Introduction 01:40 Talk begins 57:10 Q&A #MachineLearning #AI #Modularity #LearningTheory #ReinforcementLearning #LanguageModels #Algorithms #Research #TTIC