How researchers are teaching AI to learn like a child

How researchers are teaching AI to learn like a child
    Watch the video

    click to begin


    You've probably heard of machine learning.
    That's when a computer learns
    everything it needs to know from a giant dataset,
    using trial and error.
    But that's not what babies do.
    They aren't a clean slate upon entering the world.
    Babies have innate knowledge that helps them
    to voraciously learn and rapidly adapt.
    There are just some things you can't learn
    from trial and error.
    But many computer scientists argue that
    most human skills are learned
    and AI could learn them too,
    without the need for pre-loaded rules.
    Still, a growing number of researchers
    are attempting to encode AI
    with a bit of common sense.
    The current craze in AI are neural nets,
    collections of simple computing elements,
    loosely modeled on neurons in the brain,
    that adjust their connections
    as they encounter more data.
    They've produced incredible achievements
    in the past few years,
    from facial recognition to beating humans
    at poker and go.
    But neural nets require thousands of training examples
    to reliably form associations.
    And even then,
    they can produce some embarrassing blunders.
    Compare this to a child who can see an image just once
    and after that instantly recognize it in other contexts.
    Some AI's can play classic Atari games
    with super human skill,
    but when you remove all the aliens but one,
    the player inexplicably becomes a sitting duck.
    Different labs are categorizing human instincts
    and then trying to encode them into AI.
    These systems sit somewhere between
    pure machine learning and completely programmed.
    One team developed an AI called
    They've embedded the rule that:
    such a thing as objects and relationships
    between those objects exist.
    This is like a baby's innate parsing of the world
    into objects.
    In tests, once the AI learns the specific properties
    and relationships, it is able to predict the behavior
    of falling strings and bouncing balls in a box.
    Another group's "neural physics engine"
    beat less structured neural nets
    at predicting ball collisions in containers.
    And a lab created an AI which has an embedded rule
    to treat letters as objects
    and separate them from their background.
    This allowed it to solved CAPTCHAs
    better than other neural nets
    that were trained with 50,000 times more data.
    We're far away from AIs
    that can truly thinks like humans,
    But with these latest attempts
    to reproduce common sense artificially,
    researchers believe they will get closer
    to creating robots that can fully interact
    with the world the way we do.
    Machines that start like a baby and learn like a child.
    First footage of deep-sea anglerfish pair What's really behind 'gluten sensitivity'? Cheap smart wall can sense movement and electronics in a room Here's what happens when you replace toads and turtles with 3D printed replicas in the wild Watch a 'ballooning' spider take flight Mini-antibodies discovered in sharks and camels could lead to drugs for cancer and other diseases Some people have evolved bigger spleens to hunt underwater Can this robot build an IKEA chair faster than you? Scientists just 3D imaged the birthplace of stars—using sound An AI can create a 3D model of a person—from just a few seconds of video