Hacker News Reader

Comment by MichaelAza | Hacker News Reader

> "All of us know a huge number of things," said Sloan. "As babies, we crawled around and yanked on things and learned that things fall. We yanked on other things and learned that dogs and cats don't appreciate having their tails pulled."

Then let the damned thing explore! Why do we make such huge advances in robotics if not for this? A baby-like robot shouldn't be that hard to make. I'm no AI expert but couldn't we just throw our best learning AI in there and give the thing a couple of years?

Why are we trying to short-circuit human learning instead of mimicking it?

View on HN

Leaving a robot to explore and learn wouldn't make it any good. If you look at the successes and failures of AI listed in the article, modern AI is well suited for sensory recognition and pattern matching, and it's bad at cognitive reasoning in the abstract - just as it has always been. There are logical reasoners, of course, but they can only reason about whatever the programmer has put previously in the system as inference rules.

Now letting a robot roam around the premises wouldn't be much different than feeding it with slides of sensory stimuli.

The kind of experiences that the bot could only learn from moving around are beyond its sense-making; it can't gain any significant advantage from having wheels and hands and chasing the environment, because it doesn't know how to use them to enhance its knowledge and learn new kinds of things, in the way that a toddler can.

The only approach I know that seems promising in that respect is IBM's Watson - I'm not sure how it works exactly, but even if it's doing basic pattern recognition, the huge scale of parallel process and the enormous data corpus that it's feeded into Watson ''might'' just be what it takes to achieve sense-making in an emergent way.

Ya I don't like how we are biased for machines being absolutely perfect or useless. If an AI system can't do something in a few hours that takes humans years to learn or think about, it's considered a failure. Humans spend many weeks, sometimes years on things, not to mention all the time up to that point they spent learning about the world or working on similar problems.

On the other hand most algorithms will quickly plateau and not make much progress afterwards. Machine learning algorithms are nowhere near the ability of human brains and letting them run for 4 years wouldn't be much different than running them for a few weeks. (Also time on a supercomputer isn't cheap.)

Here's one team trying for that: http://www.inria.fr/en/teams/flowers

This is what I want to see happen. I want to see categorical cross-referencing events/things where round is applied to ball and ball is applied to basketballs and baseballs. I want to see orange applied to basketballs, and then a connection is made between the fruit orange and a basketball. And just let that thing work, just leave it unassisted for a year as it grows more and more connections and see where the lexical processing and visual (color and shape matching) processing takes it.

Reply on news.ycombinator.com