Stanford Seminar - Leela: A Semantic Intelligent Agent

Overview

Leela is a semantic artificially intelligent agent modeled on the theories of Jean Piaget. She builds increasingly abstract semantic models of the world from her experiences of exploration, play, and experimentation. As an agent she is able to formulate, execute, and explain her own plans.

This talk will provide an introduction to Leela's background and design and will show her in action.

About the Speaker: David Henkel-Wallace is a founder of Leela.ai, an AI research company building general semantic AI agents. His 25 year career as a technical company founder in Silicon Valley has spanned software, pharmaceuticals, and power generation. Prior to becoming an entrepreneur he was an AI researcher at research labs like Xerox PARC, MCC and the MIT AI Lab.` He is a graduate of MIT.

Syllabus

Introduction.
A quick who, where, when.
What is leela - 1/2.
What is leela - 2/2.
Shouldn't your self-driving car notice?.
Important state is not always visible.
People and systems both need "why".
Monolithic approaches don't scale.
You need abstraction, automation, and a toolkit.
Leela's semantics comes from objects and actions, not mere states There is no single "chair" perceptron or object.
Piaget's model of experimentation and construction 1. Assimilation: fit new experience into existing mental structures (schemas) 2. Accommodation: revise existing schema in light of new experience 3. Equilibrium: seek cognitive stability though assimilation and accommodation..
Power from abstraction & re-use.
Language training architecture.
Piaget's stages of development.
Leela understands instructions and can explain what she is doing. She can can explain her internal state, plans and reactions using natural language.