Learning from All Types of Experiences:
A Unifying Machine Learning Perspective
1:00PM -5:00PM (PT), Sunday August 23, 2020
@ KDD 2020
Abstract
In handling wide range of experiences ranging from data instances, knowledge, constraints, to rewards, adversaries, and lifelong interplay in an ever-growing spectrum of tasks, contemporary ML/AI research has resulted in thousands of models, learning paradigms, optimization algorithms, not mentioning countless approximation heuristics, tuning tricks, and black-box oracles, plus combinations of all above. While pushing the field forward rapidly, these results also make a comprehensive grasp of existing ML techniques more and more difficult, and make standardized, reusable, repeatable, reliable, and explainable practice and further development of ML/AI products quite costly, if possible, at all.
This tutorial will present a systematic, unified perspective of machine learning, for both a refreshing holistic understanding of the diverse learning algorithms and a guidance of operationalizing machine learning for creating problem solutions by integrating all sources of experiences.
The tutorial consists of three parts: (1) Theory: a systematic blueprint of ML that provides a unified mathematical formulation for learning with all experiences; (2) Tooling that operationalizes the framework and enables easy composition of ML solutions; (3) Computing infrastructures for productive ML, including interoperation, automatic tuning, distributing, and scheduling.
Schedule and materials
- Introduction (10min)
- Part I: Theory: The Standard Equation (90min)
A blueprint of ML paradigms for ALL experiences
Background
The Standard Equation
The zoo of optimization solvers
Applications of Standard Equation
- Part II: Tooling: Operationalizing “Learning with all experiences” (70min)
Compose your ML solutions like playing Lego
ML solution design by learning with all experiences
Compositionality in ML
Texar: an open-source tool for ML Composition
- Part III: Computing: Modern infrastructure for productive ML (40min)
Interoperation, automatic tuning, distributing, and scheduling
Interoperation: Forte
Automatic tuning: TUUN
Automatic distributing: AutoDist
Automatic scheduling: AdaptDL
Presenters
CTO @ Petuum Inc.
References
IAS Seminar on Theoretical Machine Learning: "A Blueprint of Standardized and Composable Machine Learning"
Invited Talk @ ODSC 2019: "Compositionality in Machine Learning"
AAAI 2020 Tutorial: "Modularizing Natural Language Processing"
NeurIPS 2019 Workshop: "Learning with Rich Experience (LIRE): Integration of Learning Paradigms"
ICML 2018 Workshop: "Theoretical Foundations and Applications of Deep Generative Models"