Title: The 4 pillars of machine learning: Predictions, Decisions, Analysis, Synthesis
Abstract:
I will give a unified view on “All of machine learning” from the perspective of (Bayesian) decision theory. This framing roughly follows the structure of my new book ( In particular, I identify 4 main kinds of ML: predictive modeling, decision making under uncertainty (including causality and RL), analysis (inferring latent quantities / extracting “meaning” from observed data), and synthesis (generative AI). For each one of these “pillars”, I will briefly summarize some projects that I have been involved in. Specifically, for prediction, I will discuss my NeurIPS 2023 paper on test-time adaptation for prediction under distribution shift; for sequential decision making, I will discuss my AISTATS 2022 paper on contextual bandits, and my recent COLLAS 2023 paper on a low-rank extension; for analysis, I will discuss my ICML 2020 paper on modeling sequential data using collapsed variational inference for switching non-linear dynamical systems; and for synthesis, I will discuss my ICML 2023 paper on text to image generation using masked transformers.
[ad_2]
source
