Learning About the Physical World - Comparing Humans, Deep Neural Networks, and Probabilistic Programs

Explore the foundations of human cognition and artificial intelligence in this 37-minute lecture by Josh Tenenbaum from MIT. Delve into the concept of common-sense understanding, comparing how humans, deep neural networks, and probabilistic programs perceive and interpret the physical world. Examine intuitive physics and psychology, focusing on how infants develop an understanding of object dynamics and intentional agents. Investigate the potential for engineering common-sense core knowledge using probabilistic programs and neural circuits. Gain insights into physical scene parsing, theory building through program induction, and the development of physics concepts in infants. Conclude with a comprehensive overview of how human thought is structured around basic understandings of physical objects and intentional agents, forming the core of common sense reasoning.

Intro
Al technologies
The common-sense core Human thought is structured around a basic understanding of physical objects, intentional agents, and their interactions - intuitive physics (forces, masses...) and psychology (desires, beliefs, plans...)
Common-sense scene understanding
The common sense core
Engineering the common-sense core?
Engineering the common-sense core Probabilistic programs
Intuitive physics in blocks world
(Battaglia, Hamrick, Tenenbaum, PNAS 2013) Probabilistic IPE
The intuitive physics engine (Battaglia, Hamrick. Tenenbaum, PNAS 2013)
Probabilistic dynamics in 12 month olds
Physical scene parsing
Theory building as program induction ( program-learning programs)
Development of physics in infants
Intuitive physics in the brain
Intuitive physics in neural circuits?
Conclusion The core of common sense: Human thought is structured around a basic understanding of physical objects, intentional agents, and their interactions