It’s been another year of relentless artificial-intelligence
hype and incremental AI achievement. Machines still beat humans only in
carefully constructed environments or at narrow tasks.
The good news is that, as the technology progresses, the
race for leadership is still wide open, and even Europe, where politicians
fret that the continent is lagging behind China and the US, is still quite
competitive.
According to the Artificial Intelligence Index 2018 annual
report, whose steering committee includes leading AI scholars such as Yoav
Shoham of Stanford University and Erik Brynjolfsson of the Massachusetts
Institute of Technology, AI has progressed on all the measures tracked.
Some of the metrics, from the number of published papers and
conference attendance, to mentions on corporate earnings calls and in
parliamentary hearings, measure the hype. Others reflect improving performance.
This year, AI has become more accurate and much faster at
image detection. It’s also improved at parsing the grammatical structure
of sentences, answering multiple-choice questions and translation. Whether
this progress brings us closer to truly superhuman AI is a different
matter.
On the translation front, a measure called Bilingual
Evaluation Understudy is used to determine accuracy. It compares machine-translated
sentences to those rendered by human experts, and this year almost half the
machine translations between English and German news articles measured up to
the human ones.
This year, Microsoft announced with much fanfare that
its AI did just as well as humans in translating news from Chinese into
English. But the underlying paper reports much lower scores for the
Chinese translations than for the separately published German ones, and
accuracy scores from human evaluators of between 50% and 70%.
Machine-translation algorithms still produce plenty of gibberish and are really
mostly useful, in a limited way, to humans with some understanding of both
languages and the context.
Improved image-recognition has worked wonders in some fields
of medicine. For example, Google has developed a system for grading prostate
cancer that does it more accurately than US pathologists, and a Stanford team
has achieved similar success with skin cancer. Where lots of data exist and
precision is valued, AI can help humans make better decisions, even though it
still messes up regularly when trained on biased data sets or is intentionally
tricked. Humans are less prone to misidentifying objects and are better
able to correct for their biases.
Almost human
Data-mining and question-answering skills can make AI appear
almost human at times. This year, IBM presented the current iteration of its
Project Debater, which tries to debate humans hewing to the rules of such
competitions. The exercise looks impressive — the machine instantaneously
gathers and orders information, packs it into grammatically correct
sentences and inserts pre-written jokes almost in the correct places. But as an
AI expert who was present discovered, it tended merely to repeat its points in
response to arguments. While the idea of having a machine, with its
superhuman ability to analyse data, take part in brainstorms is
exciting, "We are most certainly not on the verge of seeing AI
systems out-debating their human counterparts," wrote the expert, Chris
Reed of the University of Dundee in Scotland. "Today’s AI technology is as
far from these scenarios as the Romans’ experiments with steam power were from
the industrial revolution," he concluded.
As often happens with technological advancement, AI gets too
much attention too early. But if in previous years some AI scholars grumbled
that the hype might impede progress because people would become disappointed in
the unfulfilled promise of a shiny toy, attention to AI has become too
sustained and the financial and intellectual resources thrown at it too
enormous for that to happen. Now, competing in AI is a matter of prestige for
major nations.
So far, the US, Europe and China all have their
strengths. Data in the Artificial Intelligence Index report show the US
as the runaway leader in patents; along with China, it leads in the number of
papers submitted to and accepted by major AI conferences. But it is in Europe
where the greatest number of AI papers (28% of the total, compared with 25
% for China and 17 % for the US) are published. A report published by the
European Commission’s Joint Research Centre this month says that the European
Union is home to a quarter of the approximately 35 000 entities working in
artificial intelligence today, compared with 28% for the US and 23% for China.
According to McKinsey & Co., Europe also matches
competitors when it comes to AI adoption in business, especially in
process automation.
This is likely to come as a surprise to European
leaders, especially German and French ones, who often talk about falling
behind. Earlier this month, German Economy Minister Peter Altmaier supported
the idea of a pan-European state-led corporation, along the lines of Airbus, to
compete in AI.
Europe doesn’t really need massive state interference to
catch up, as it did in the 1960s and 1970s when Boeing dominated the aircraft
industry. But the EU and governments in North America and China will be pouring
more resources into AI in the coming years, and distinct development models are
likely to crystallise in the key competing countries as regulation follows the
money. The Joint Research Centre report names three approaches that are easy to
match to their regions of origin: "AI for profit," "AI for
control" and "AI for society," a discipline it defines as "a
human-centred, ethical and secure approach."
Regardless of how well the technology will eventually work,
major nations have already co-opted it for soft power and ideological
competition. It’s a rerun of last century’s space race, not seen in this pure a
form for many decades.