Explore the intersection of physics and artificial intelligence in this comprehensive conference talk from KDD 2020. Delve into the four paradigms of scientific discovery, examining the interplay between theory and data. Investigate the limitations of big data approaches and the importance of generalization in both physics and AI. Analyze computational complexity classes and their relevance to AI and physics problems. Discover physics-informed neural networks (PINN) and physics-guided neural networks (PGNN), and their applications in explainable AI. Examine open questions in neural networks, including the potential for a statistical physics theory of deep learning. Gain insights into information bottlenecks and their role in both physics and neural networks. Enhance your understanding of the evolving relationship between physics and artificial intelligence through this in-depth presentation.
KDD 2020: Physics Inspired Models in Artificial Intelligence
Association for Computing Machinery (ACM) via YouTube
Overview
Syllabus
Intro
Terminology
Motivation: Why Physics & AI?
Why this Tutorial?
Tutorial Goals
Interplay of Physics and Al
The Four Paradigms
Theory vs. Data?
Limitations of the 4th Paradigm
Cautionary Tale: Problems with Big Data
Parameters Galore!
Physics: Tycho Brahe to Kepler to Newton
A Brief History of Physics & Al
Generalization in Physics & Al
Generalization in Neural Nets
Generalization: Observations
Computational Complexity, Al & Physics
Complexity Classes
3-SAT and Phase Transitions
Problems: Complexity
Interpretability & Explainability in Al/ML
Properties of XAI
Physics Informed Neural Nets (PINN)
Physics-guided Neural Network (PGNN)
Physics & Explainable Al: An Illustration
Results Summary
Open Questions in Neural Networks
Statistical physics theory of Deep Learning?
Information Bottleneck & Neural Nets
Information Bottlenecks & Physics
The Committee Machine
Taught by
Association for Computing Machinery (ACM)
