know to drive cautiously when the car in front of
us is swerving dangerously, indicating that the
driver is either drunk, sleepy, or distracted. And
we know slippery roads are dangerous, even if it’s
RXU¿UVWWLPHGULYLQJRQRQH
But current AI systems lack common sense,
intuitive physics, basic psychology, and other
elements that make us humans general problem
solvers. Deep-learning systems are number-
crunching engines that compare the data they’re
seeing with what they’ve seen before.
“At present, I would say that humans are still far
more trustworthy drivers than self-driving cars if
those cars are going to be allowed to drive ‘under
all conditions,’” Mitchell says.
In recent years, there has been a push to integrate
common sense into deep-learning systems. In last
year’s NeurIPS AI conference, Yoshua Bengio, an
award-winning deep-learning pioneer, spoke of
³V\VWHPGHHSOHDUQLQJ ́DOJRULWKPVWKDWSDUVH
data into the kind of abstract, high-level concepts
that are associated with human intelligence.
Researchers at MIT and IBM are working on
hybrid AI systems that integrate human-encoded
rules into neural networks. Other scientists are
working on self-supervised learning, a type of
deep learning that, like a child, can explore its
environment and discover the rules by itself.
%XWDOOWKHVHH̆RUWVDUHVWLOOLQWKHSUHOLPLQDU\
research stage and haven’t resulted in any
breakthroughs. If the past is prologue, it’ll be
years before any of these projects turn into
commercially viable products.
B
e
n
D
i
c
k
s
o
n
Human drivers
don’t need to
know every-
thing to make
reasonable de-
cisions when
facing new
situations.
