Page 144 - ITU Journal - ICT Discoveries - Volume 1, No. 2, December 2018 - Second special issue on Data for Good

,78 -2851$/ ,&7 'LVFRYHULHV 9RO 'HFHPEHU

simulations. Before getting good at its job, the the algorithm encounters cases it has not seen
machine often ‘got it’ wrong, and was told so: it before (a dog with a flat face or a human with darker
(virtually) crashed into trees, crushed people, and skin than in the data set it was trained on)?
fled when seeing the police; all things considered Fundamentally, how should those estimations,
bad. Through these trials, errors, and feedbacks, it predictions, and prescriptions be used, and by
started being able to drive autonomously. Another whom, when, and if at all?
machine looked at the picture of a cat and, when
prompted, concluded it was a dog. It was told These risks are real. They need to be known and
“wrong!” and asked to try again with different addressed to limit the worst typical side effects of
photographs, many times over. technological change, at least in the short run,
including widening inequities. But big data and AIs
Through these iterations, these machines learned are neither ‘black magic’; nor are the algorithms
what features and combination of features of what running them complete ‘black boxes’. Given their
they were seeing were most systematically ubiquity and power, it is important to understand
associated with the right result. The algorithm, the how they work and what insights we could glean
series of steps classifying, organizing, ranking from them to promote positive social change.
information and tasked with concluding “cat!” or Critically, it is not (just) about using AI to optimize
“dog!” figured out that the longer the nose, the more supply chains (and more), which will continue to
likely the “thing” was to be “dog”, whereas have major impacts on societies and economies, but
considering whether it had long or short hair was about being inspired and supported by AI to improve
not a very valuable use of its neurons. It was human systems.
learning how to “connect the dots”. The machine
was learning. The gist of big data and current AI(s) What is the ‘good magic’ of current AIs? In short, the
is machine(s) learning. good magic, is its “credit assignment (or reward)
function”. It is the ability to assign credit for what
Of course, there are many more caveats and “works”; in other words what allows an algorithm
complexities than these, but for most intents and to get the right (intended) result. In the example
purposes it suffices to understand that current above, the computer tasked with telling a dog from
‘narrow’ AI (as opposed to a ‘general’ AI that fuels a cat will extract millions of features from the image
the most vivid fears about robots taking over the it sees, then assemble them in millions of ways, take
world, which does not seem like a realistic outcome guesses, and over time, learn which combinations of
in the foreseeable future) is about this: getting lots paths allow it to get the right answer (assuming
of data as inputs and learning how to connect them everyone “calls a cat a cat”, as the French say )
1
to output data in the form of desirable or observed almost all the time. The reward function and ability
outcomes. Through training, testing, and learning to learn through iterations lead to reinforcement of
based on past cases, the machine is able to land on the combination of features to look for and use. In
the “right results”. contrast, those that lead to the wrong result will be
weakened. The machine will grow an incentive to
The applications and implications of this are not use them.
already far-reaching. Is this person going to like this
book because someone just like him or her As it turns out, or so we think, applying the core
(including him or her last month) did? Is this principles and requirements of AI to entire human
teenager on the verge of dropping out of school? Is systems in a consistent, careful manner to design
this person Kieran McKay or Abigail Adeyemi? and deploy “human-machine (eco)systems” could
Should he or she get a loan? Should the driverless be quite transformative, for the better.
car kill a pregnant woman or five elderly people if it
has no choice but to run over either? Several tough
related questions come to mind and fuel ongoing
debates. If algorithms seem racist, is it because their
developers embed their biases or rather because
predictions repeat past biases? What happens when

1 From the French phrase “Appeler un chat un chat” which
means “Calling a spider, a spider”

‹ ,QWHUQDWLRQDO 7HOHFRPPXQLFDWLRQ 8QLRQ