For all the remarkable recent progress in AI research, we are still very far from creating machines that think and learn as well as people do. As Meta AI’s Chief AI Scientist Yann LeCun notes, a teenager who has never sat behind a steering wheel can learn to drive in about 20 hours, while the best autonomous driving systems today need millions or billions of pieces of labeled training data and millions of reinforcement learning trials in virtual environments. And even then, they fall short of human’s ability to drive a car reliably.
What will it take to build AI that approaches human-level capabilities? Is it simply a matter of more data and bigger AI models?
As part of Meta AI’s Inside the Lab event on February 23, 2022, LeCun is sketching an alternate vision for building human-level AI. LeCun proposes that the ability to learn “world models” — internal models of how the world works — may be the key.
Meta AI is sharing some of LeCun’s ideas in brief here, including his proposal for a modular, configurable architecture for autonomous intelligence, as well as key challenges the AI research community must address to build such a system. We typically share the results of our research — by publishing papers, code, and data sets, as well as blog posts — when they are completed. But in keeping with Meta AI’s open-science approach, we are taking this opportunity to preview our research vision and ideas in the hope that it spurs discussion and collaboration among AI researchers. The simple fact is that we will need to work together to solve these extraordinarily challenging, exciting problems.
We plan to share more details on LeCun’s vision in an upcoming position paper.
“Human and nonhuman animals seem able to learn enormous amounts of background knowledge about how the world works through observation and through an incomprehensibly small amount of interactions in a task-independent, unsupervised way,” LeCun says. “It can be hypothesized that this accumulated knowledge may constitute the basis for what is often called Common sense.”
And Common sense can be seen as a collection of models of the world that can guide on what is likely, what is plausible, and what is impossible.
This allows humans to plan effectively in unfamiliar situations. That teen driver may not have driven over snow before, for example, but he (hopefully) knows that snow can be slippery and send his car into a skid he drives too aggressively.
Common sense knowledge allows animals not just to predict future outcomes but also to fill in missing information, whether temporally or spatially. When a driver hears the sound of metal smashing together nearby, he knows immediately that there’s been an accident — even without seeing the vehicles involved.
The idea that humans, animals, and intelligent systems use world models goes back many decades in psychology and in fields of engineering such as control and robotics. LeCun proposes that one of the most important challenges in AI today is devising learning paradigms and architectures that would allow machines to learn world models in a self-supervised fashion and then use those models to predict, reason, and plan. His outline regroups ideas that have been proposed in various disciplines, such as cognitive science, systems neuroscience, optimal control, reinforcement learning, and “traditional” AI, and combines them with new concepts in machine learning, such as self-supervised learning and joint-embedding architectures.
LeCun proposes an architecture composed of six separate modules. Each is assumed to be differentiable, in that it can easily compute gradient estimates of some objective function with respect to its own input and propagate the gradient information to upstream modules.
