Artificial General Intelligence

Columbia University - Summer A 2023

Class is held on Monday and Wednesday 5:30-8:40pm

Course staff and office hours

Instructor: Prof. Iddo Drori

TA: TBD

The Artificial General Intelligence course includes three parts: (1) Consciousness, (2) Human Intelligence, and (3) Artificial General Intelligence.

The first part covers the elements required for a conscious neural network. Consciousness does not mean intelligence, does not require biology, and only requires combining existing machine learning and AI capabilities covered in the first part of the class: (1) self-report, (2) conversational ability available in language models, (3) domain-general abilities available in a generalist agent, (4) sensory processing available in vision-language transformers, (5) the ability to act available in vision-language-action transformers, (6) world models, (7) self models, (8) recurrent processing available in sequence models, (9) a global workspace, and (10) a unified agency.

The second part of the class covers human intelligence from a neuroscience perspective, including synaptic transmission, perception, movement, emotion, and motivation. We then describe human development and emergent behavior, followed by the neural mechanisms of learning, memory, language, and cognition.

The third part presents a path toward artificial general intelligence including quality-diversity, and open ended exploration. The class covers recent work published in Nature, Science, PNAS, ICLR, NeurIPS, ICML, and AAAI, and is based on notes for a book in progress titled Artificial General Intelligence by Iddo Drori, commissioned by Cambridge University Press.

Each lecture is three and a half hours long with two ten minute breaks. During the third hour, students present papers in a seminar style class.

Lecture 1: Introduction

Part I: Consciousness

Lecture 2: Vision, language, and action transformers

Lecture 3: Domain general models, generalist agents

Lecture 4: Recurrent models, world and self-models

Lecture 5: Global workspace models, memory, and unified agency

Lecture 6: A conscious neural network

Part II: Human Intelligence

Lecture 7: Synaptic transmission and perception, movement, emotion

Lecture 8: Development and emergence of behaviour

Lecture 9: Learning, memory, language and cognition

Part III: Artificial General Intelligence

Lecture 10: Foundation models, reinforcement learning, and generative AI

Lecture 11: Quality-diversity, open ended exploration

Lecture 12: Meta, multi-task, few-shot, continual, and lifelong learning