In the past few years, deep-learning algorithms have gradually become
better at detecting objects in images. But one problem remains: deep-
learning systems become unreliable when encountering new situations.
“Today’s state-of-the-art self-driving cars can perform autonomously
quite well as long as their ‘conditions’ remain predictable and well
within the kinds of driving situations they have been trained on,”
Mitchell says. “The problem comes with unexpected situations that don’t
closely match the cars’ training regimes.”
,Q7HVOD¶V$,DOJRULWKPIDLOHGWRGHWHFWDZKLWHWUDFWRUWUDLOHUWUXFN
against a bright sky, which resulted in a fatal crash because the car’s AI
had not seen such scenery during its training and was unable to detect
the truck.
One solution is what’s become known as “throwing more data at the
problem”: Train the deep-learning algorithm with more and more data
in hopes of covering every possible situation. This is one area in which
Tesla excels, because it is constantly collecting data from its hundreds of
WKRXVDQGVRIFDUVWRWUDLQDQG¿QHWXQHLWVDOJRULWKPV
But while Tesla cars will probably never mishandle “white tractor
against a bright sky” and “concrete barrier in the middle of the road”
VLWXDWLRQVDJDLQVXFKHGJHFDVHV²DVUDUHLQFLGHQWVDUHNQRZQ²DUHWRR
numerous and can’t be predicted in advance.
“People in the self-driving-car industry talk about the ‘long tail’ of
XQOLNHO\VLWXDWLRQV²RQHVWKDWGRQRWQRUPDOO\DSSHDULQGD\WRGD\
driving, so are unlikely for any given vehicle to encounter, but are so
numerous that each ‘tail’ situation will occur for some car somewhere on
a regular basis,” Mitchell says. “It’s impossible to train an AI system on
all such situations.”
AI NEEDS SOME COMMON SENSE
Human drivers don’t need to know everything to make reasonable
decisions when facing new situations. We handle edge cases by tapping
into our knowledge of the world and people. We know that an
unattended child might risk running in the middle of the street. We
B
e
n
D
i
c
k
s
o
n
