Cover of "The Book of Why"
A lot of artificial intelligence researchers have put a pause on their quest for a super-intelligent machine, frustrated by the lack of recent progress. But among those still in the hunt are Judea Pearl, winner of the Turing Award, the highest prize in computer science, and author of "The Book of Why," in which he proposes a new map to intelligent machines.
- Pearl calls his approach "causal reasoning," the ability to infer the whys and hows of a situation.
- While it won't get to artificial general intelligence itself, causal reasoning is the eventual route there, and thus will mark a "mini-revolution," he tells Axios.
The problem with machine learning, Pearl said, is that it rests on correlation and association, which have made remarkable achievements but are still elementary in terms of true thinking — and they can only take the field so far.
- He told me the story of an AI program watching a rooster crow day after day before the sun rises, and determining that it causes the sun to rise.
- That is an example of correlation. But, Pearl writes in the book, "Causal explanations, not dry facts, make up the bulk of our knowledge, and should be the cornerstone of machine intelligence."
- The goal of AI researchers "is to produce machines with humanlike intelligence, able to converse with and guide humans. Deep learning has instead given us machines with truly impressive abilities but no intelligence."
Causal reasoning differs by working through a situation without specific training, said Pearl, and borrows from methods already in use by social and health science researchers.
- For instance, a self-driving car equipped with causal reasoning could encounter a situation for which it has no data, and instantly adapt, said Pearl.
- "This is the route [to general artificial intelligence] and these are the building blocks," he said. "On top of them, we need to build capabilities such as: social intelligence, humor, and perhaps even fear of death."
- "The mini-revolution I am predicting will liberate machine learning from its current predicaments of opaqueness, forgetfulness and lack of explainability," he said. "Plus it will allow machine to learn answers to questions we really care about, (e.g., what caused this accident?) not merely associations."
Go deeper: In this Bloomberg video, Facebook's Yann Lecun discusses how to advance AI.